{ "version": 2, "completed": true, "production_settings_changed": false, "stages": [ { "stage": 1, "measured": true, "statistics": { "echo_on": { "n": 6, "wer": 0.6111111111111112, "cer": 0.5959595959595959, "false_finals": 0, "full_refusals": 0, "final_median_ms": 768.0222541093826, "minimum_final_ms": 658.028107136488, "maximum_final_ms": 837.9631489515305, "capture_gaps": 0 }, "echo_off": { "n": 6, "wer": 0.7407407407407407, "cer": 0.7104377104377104, "false_finals": 6, "full_refusals": 1, "final_median_ms": 666.1063972860575, "minimum_final_ms": 644.3984396755695, "maximum_final_ms": 796.7720404267311, "capture_gaps": 0 } }, "selection": { "profile": { "noise": "native", "echo": true, "regulator": "none", "normalization": "baseline", "silence": 450, "partial": 500, "mode": "online", "agc": false }, "selected": "echo_on", "totals": { "echo_off": { "errors": 40, "reference_units": 54, "false": 6, "failures": 0, "latencies": [ 666.9420152902603, 666.1063972860575, 658.2467276602983, 796.7720404267311, 644.3984396755695 ] }, "echo_on": { "errors": 33, "reference_units": 54, "false": 0, "failures": 0, "latencies": [ 658.028107136488, 837.9631489515305, 768.0222541093826 ] } }, "rejected": { "echo_off": [ "new_false_finals", "new_splits_or_finals", "new_full_refusal", "numbers_or_codes_degraded" ] }, "acceptance": "declared_criteria_screening_complete; held_out_complete; production_unchanged" } }, { "stage": 2, "measured": true, "statistics": { "rms24": { "n": 6, "wer": 0.35960591133004927, "cer": 0.21390374331550802, "false_finals": 0, "full_refusals": 0, "final_median_ms": 671.9748312607408, "minimum_final_ms": 668.3018133044243, "maximum_final_ms": 752.3663267493248, "capture_gaps": 0 }, "baseline": { "n": 6, "wer": 0.37438423645320196, "cer": 0.23071046600458364, "false_finals": 0, "full_refusals": 0, "final_median_ms": 676.2350238859653, "minimum_final_ms": 666.0593934357166, "maximum_final_ms": 703.7342861294746, "capture_gaps": 0 }, "rms22": { "n": 6, "wer": 0.37438423645320196, "cer": 0.226890756302521, "false_finals": 0, "full_refusals": 0, "final_median_ms": 679.3657364323735, "minimum_final_ms": 672.1029374748468, "maximum_final_ms": 792.5739027559757, "capture_gaps": 0 } }, "selection": { "profile": { "noise": "native", "echo": true, "regulator": "none", "normalization": "baseline", "silence": 450, "partial": 500, "mode": "online" }, "selected": "baseline", "totals": { "baseline": { "errors": 76, "reference_units": 203, "false": 0, "failures": 0, "latencies": [ 672.9609686881304, 674.4783017784357, 666.0593934357166, 677.991745993495, 703.7342861294746, 688.5090842843056 ] }, "rms24": { "errors": 73, "reference_units": 203, "false": 0, "failures": 0, "latencies": [ 670.9083989262581, 669.4748643785715, 752.3663267493248, 673.0412635952234, 674.0030236542225, 668.3018133044243 ] }, "rms22": { "errors": 76, "reference_units": 203, "false": 0, "failures": 0, "latencies": [ 679.1905891150236, 674.9217323958874, 792.5739027559757, 679.5408837497234, 683.2816433161497, 672.1029374748468 ] } }, "rejected": { "rms24": [ "quiet_improvement_below_20_percent" ], "rms22": [ "quiet_improvement_below_20_percent" ] }, "acceptance": "declared_criteria_screening_complete; held_out_complete; production_unchanged" } }, { "stage": 3, "measured": true, "statistics": { "none": { "n": 8, "wer": 0.8198198198198198, "cer": 0.6825068870523416, "false_finals": 0, "full_refusals": 1, "final_median_ms": 683.1254921853542, "minimum_final_ms": 647.7418467402458, "maximum_final_ms": 718.667533248663, "capture_gaps": 0 }, "native": { "n": 8, "wer": 0.8378378378378378, "cer": 0.7155647382920111, "false_finals": 0, "full_refusals": 2, "final_median_ms": 670.5093719065189, "minimum_final_ms": 648.7707663327456, "maximum_final_ms": 713.8287015259266, "capture_gaps": 0 }, "rnnoise": { "n": 8, "wer": 0.8603603603603603, "cer": 0.7334710743801653, "false_finals": 0, "full_refusals": 1, "final_median_ms": 668.0963970720768, "minimum_final_ms": 638.4422443807125, "maximum_final_ms": 730.3509339690208, "capture_gaps": 0 } }, "selection": { "profile": { "noise": "native", "echo": true, "regulator": "none", "normalization": "baseline", "silence": 450, "partial": 500, "mode": "online", "agc": false }, "selected": "native", "totals": { "none": { "errors": 182, "reference_units": 222, "false": 0, "failures": 1, "decode_count": 130, "latencies": [ 683.1254921853542, 647.7418467402458, 689.4971318542957, 718.667533248663, 674.9691143631935 ] }, "native": { "errors": 186, "reference_units": 222, "false": 0, "failures": 1, "decode_count": 117, "latencies": [ 677.8706982731819, 663.148045539856, 648.7707663327456, 713.8287015259266 ] }, "rnnoise": { "errors": 191, "reference_units": 222, "false": 0, "failures": 1, "decode_count": 155, "latencies": [ 730.3509339690208, 638.4422443807125, 663.488581776619, 676.8198870122433, 668.0963970720768 ] } }, "rejected": { "none": [ "new_splits_or_finals", "clean_codes_or_endings_degraded" ], "rnnoise": [ "new_splits_or_finals", "clean_codes_or_endings_degraded", "no_noise_wer_improvement" ] }, "acceptance": "declared_criteria_screening_complete; held_out_complete; production_unchanged", "tie_break": "quality first; stage 6 minimizes decode within accepted latency bound; other stages minimize final latency" } }, { "stage": 4, "measured": true, "statistics": { "fixed": { "n": 6, "wer": 0.44594594594594594, "cer": 0.2610192837465565, "false_finals": 0, "full_refusals": 0, "final_median_ms": 690.3958087787032, "minimum_final_ms": 666.6546799242496, "maximum_final_ms": 720.8775877952576, "capture_gaps": 0 }, "causal": { "n": 6, "wer": 0.5225225225225225, "cer": 0.3023415977961432, "false_finals": 0, "full_refusals": 0, "final_median_ms": 696.277148090303, "minimum_final_ms": 672.9392018169165, "maximum_final_ms": 753.5087279975414, "capture_gaps": 0 }, "native_agc": { "n": 6, "wer": 0.42342342342342343, "cer": 0.24793388429752067, "false_finals": 0, "full_refusals": 0, "final_median_ms": 676.6237840056419, "minimum_final_ms": 663.9434397220612, "maximum_final_ms": 686.1139684915543, "capture_gaps": 0 } }, "selection": { "profile": { "noise": "native", "echo": true, "regulator": "none", "normalization": "baseline", "silence": 450, "partial": 500, "mode": "online", "agc": true }, "selected": "native_agc", "totals": { "fixed": { "errors": 99, "reference_units": 222, "false": 0, "failures": 0, "decode_count": 177, "latencies": [ 692.9315533488989, 713.3595868945122, 720.8775877952576, 687.8600642085075, 666.6546799242496, 686.2071584910154 ] }, "native_agc": { "errors": 94, "reference_units": 222, "false": 0, "failures": 0, "decode_count": 178, "latencies": [ 677.9460795223713, 663.9434397220612, 686.1139684915543, 676.5507906675339, 673.8072354346514, 676.69677734375 ] }, "causal": { "errors": 116, "reference_units": 222, "false": 0, "failures": 0, "decode_count": 177, "latencies": [ 695.0124949216843, 694.463200867176, 753.5087279975414, 672.9392018169165, 699.5053719729185, 697.5418012589216 ] } }, "rejected": { "fixed": [ "new_splits_or_finals", "clean_codes_or_endings_degraded" ], "causal": [ "clean_codes_or_endings_degraded", "source_pause_amplification" ] }, "acceptance": "declared_criteria_screening_complete; held_out_complete; production_unchanged", "tie_break": "quality first; stage 6 minimizes decode within accepted latency bound; other stages minimize final latency" } }, { "stage": 5, "measured": true, "statistics": { "300": { "n": 8, "wer": 0.21019108280254778, "cer": 0.16301969365426697, "false_finals": 0, "full_refusals": 0, "final_median_ms": 499.6019033715129, "minimum_final_ms": 490.2112130075693, "maximum_final_ms": 516.661811619997, "capture_gaps": 0 }, "250": { "n": 8, "wer": 0.21019108280254778, "cer": 0.16301969365426697, "false_finals": 0, "full_refusals": 0, "final_median_ms": 440.90111274272203, "minimum_final_ms": 418.3173570781946, "maximum_final_ms": 452.54023745656013, "capture_gaps": 0 }, "450": { "n": 8, "wer": 0.20382165605095542, "cer": 0.1739606126914661, "false_finals": 0, "full_refusals": 0, "final_median_ms": 671.3884519413114, "minimum_final_ms": 655.7007934898138, "maximum_final_ms": 792.8117513656616, "capture_gaps": 0 } }, "selection": { "profile": { "noise": "native", "echo": true, "regulator": "none", "normalization": "baseline", "silence": 450, "partial": 500, "mode": "online" }, "selected": "450", "totals": { "450": { "errors": 32, "reference_units": 157, "false": 0, "failures": 0, "decode_count": 134, "latencies": [ 663.902048021555, 792.8117513656616, 658.1534706056118, 699.8717784881592, 692.9726041853428, 655.7007934898138, 669.1605169326067, 673.616386950016 ] }, "300": { "errors": 33, "reference_units": 157, "false": 0, "failures": 0, "decode_count": 132, "latencies": [ 503.8359183818102, 516.661811619997, 501.153452321887, 495.5690838396549, 496.38419039547443, 498.05035442113876, 506.0132946819067, 490.2112130075693 ] }, "250": { "errors": 33, "reference_units": 157, "false": 0, "failures": 0, "decode_count": 131, "latencies": [ 439.81083296239376, 418.3173570781946, 452.54023745656013, 445.79111225903034, 433.42816457152367, 446.22885808348656, 433.9854568243027, 441.9913925230503 ] } }, "rejected": { "300": [ "new_splits_or_finals", "new_phrase_splits", "question_split_against_declared_reference" ], "250": [ "new_splits_or_finals", "new_phrase_splits", "question_split_against_declared_reference" ] }, "acceptance": "declared_criteria_screening_complete; held_out_complete; production_unchanged", "tie_break": "quality first; stage 6 minimizes decode within accepted latency bound; other stages minimize final latency" } }, { "stage": 6, "measured": true, "statistics": { "800": { "n": 7, "wer": 0.28776978417266186, "cer": 0.15497737556561086, "false_finals": 0, "full_refusals": 0, "final_median_ms": 666.1147885024548, "minimum_final_ms": 646.8217819929123, "maximum_final_ms": 727.2902000695467, "capture_gaps": 0 }, "500": { "n": 7, "wer": 0.2517985611510791, "cer": 0.15497737556561086, "false_finals": 0, "full_refusals": 0, "final_median_ms": 661.8186123669147, "minimum_final_ms": 645.9328476339579, "maximum_final_ms": 680.2900172770023, "capture_gaps": 0 }, "no_taps": { "n": 1, "wer": 0.5555555555555556, "cer": 0.5252525252525253, "false_finals": 0, "full_refusals": 0, "final_median_ms": 784.0333878993988, "minimum_final_ms": 784.0333878993988, "maximum_final_ms": 784.0333878993988, "capture_gaps": 0 }, "1000": { "n": 7, "wer": 0.26618705035971224, "cer": 0.16289592760180996, "false_finals": 0, "full_refusals": 0, "final_median_ms": 680.3294643759727, "minimum_final_ms": 648.8161832094193, "maximum_final_ms": 770.5769315361977, "capture_gaps": 0 }, "taps": { "n": 1, "wer": 0.5555555555555556, "cer": 0.5252525252525253, "false_finals": 0, "full_refusals": 0, "final_median_ms": 708.8331952691078, "minimum_final_ms": 708.8331952691078, "maximum_final_ms": 708.8331952691078, "capture_gaps": 0 } }, "selection": { "profile": { "noise": "native", "echo": true, "regulator": "none", "normalization": "baseline", "silence": 450, "partial": 500, "mode": "online" }, "selected": "500", "totals": { "500": { "errors": 35, "reference_units": 139, "false": 0, "failures": 0, "decode_count": 118, "latencies": [ 678.3836726099253, 680.2900172770023, 645.9328476339579, 661.8186123669147, 646.9500716775656 ] }, "800": { "errors": 40, "reference_units": 139, "false": 0, "failures": 0, "decode_count": 86, "latencies": [ 662.5798381865025, 669.8881089687347, 646.8217819929123, 666.1147885024548, 727.2902000695467 ] }, "1000": { "errors": 37, "reference_units": 139, "false": 0, "failures": 0, "decode_count": 79, "latencies": [ 680.3294643759727, 671.9141770154238, 770.5769315361977, 718.6523992568254, 648.8161832094193 ] } }, "rejected": { "800": [ "clean_codes_or_endings_degraded" ], "1000": [ "clean_codes_or_endings_degraded" ] }, "acceptance": "declared_criteria_screening_complete; held_out_complete; production_unchanged", "tie_break": "quality first; stage 6 minimizes decode within accepted latency bound; other stages minimize final latency" } }, { "stage": 7, "measured": true, "statistics": { "combined": { "n": 18, "wer": 0.6213151927437641, "cer": 0.5426356589147286, "false_finals": 0, "full_refusals": 6, "final_median_ms": 680.5342324078083, "minimum_final_ms": 647.6319245994091, "maximum_final_ms": 754.2067319154739, "capture_gaps": 0 }, "baseline": { "n": 18, "wer": 0.5260770975056689, "cer": 0.416582406471183, "false_finals": 0, "full_refusals": 4, "final_median_ms": 659.8912691697478, "minimum_final_ms": 632.633225992322, "maximum_final_ms": 705.2855379879475, "capture_gaps": 0 } }, "selection": { "accepted": false, "reasons": [ "held_out_wer_regression", "lost_ending", "new_full_refusal", "new_terminal_asr_error" ], "recommendation": "baseline", "production_enabled": false, "baseline_samples": 18, "combined_samples": 18, "identical_requested_configuration": false, "identical_effective_configuration": true, "effective_comparison": "AGC uses the actual getSettings value; other declared processing settings are unchanged", "interpretation": "paired held-out guard, not evidence of a statistically significant improvement" } } ], "weaknesses": [ { "run": "id-0ce1f63602a83d", "issue": "full_refusal", "stage": 3, "profile": "none", "measured_wer": 1.0, "reasons": [ { "owner_ms": 8678.051892668009, "version": 1, "result_type": "error", "session_id": "id-d131543cbb7b00", "stream_id": "id-a40a3d1ba298a4", "phrase_id": "id-4bbf6d988bfacc", "revision": 10, "text": "", "code": "audio_queue_overflow", "traceparent": "00-b57820505ed642b829d69c87ed896ecb-835cb9e88bbb03e0-01", "timings": { "version": 1, "asr": { "backend": "streaming_worker", "worker_queue_ms": 0.6772223860025406, "worker_decode_ms": 1877.8796307742596, "decode_count": 11, "otel_trace_id": "b57820505ed642b829d69c87ed896ecb", "otel_span_id": "4102df135bf16d23", "last_speech_to_final_ms": 9.238358587026596 } } }, { "owner_ms": 8692.292790859938, "session_id": "id-d131543cbb7b00", "stream_id": "id-a40a3d1ba298a4", "phrase_id": "id-4bbf6d988bfacc", "version": 1, "result_type": "error", "revision": 1000000, "text": "", "timings": { "version": 1, "asr": { "backend": "streaming_worker" } }, "code": "bridge_failure" } ] }, { "run": "id-0ce1f63602a83d", "issue": "ending_not_exact", "stage": 3, "profile": "none" }, { "run": "id-ad2886a285b230", "issue": "ending_not_exact", "stage": 3, "profile": "native" }, { "run": "id-a21fb20760517d", "issue": "ending_not_exact", "stage": 5, "profile": "300" }, { "run": "id-d09485f285d150", "issue": "ending_not_exact", "stage": 5, "profile": "450" }, { "run": "id-2edab661426ddd", "issue": "ending_not_exact", "stage": 3, "profile": "rnnoise" }, { "run": "id-6abed89bb80ccd", "issue": "full_refusal", "stage": 7, "profile": "combined", "measured_wer": 1.0, "reasons": [ { "owner_ms": 11462.86642178893, "version": 1, "result_type": "error", "session_id": "id-d131543cbb7b00", "stream_id": "id-099c0fb4e5d7b8", "phrase_id": "id-b7f4aafc0b7edc", "revision": 15, "text": "", "code": "audio_queue_overflow", "traceparent": "00-ff30217185ee2d3fe30e0513e7b7ca58-0d08b6dd0ccb113a-01", "timings": { "version": 1, "asr": { "backend": "streaming_worker", "worker_queue_ms": 0.8784458041191101, "worker_decode_ms": 2475.3578677773476, "decode_count": 14, "otel_trace_id": "ff30217185ee2d3fe30e0513e7b7ca58", "otel_span_id": "787e9063415dd390", "last_speech_to_final_ms": 6.66138157248497 } } }, { "owner_ms": 11501.869019120932, "session_id": "id-d131543cbb7b00", "stream_id": "id-099c0fb4e5d7b8", "phrase_id": "id-b7f4aafc0b7edc", "version": 1, "result_type": "error", "revision": 1000000, "text": "", "timings": { "version": 1, "asr": { "backend": "streaming_worker" } }, "code": "bridge_failure" } ] }, { "run": "id-6abed89bb80ccd", "issue": "ending_not_exact", "stage": 7, "profile": "combined" }, { "run": "id-7615cee3acdfc8", "issue": "ending_not_exact", "stage": 7, "profile": "baseline" }, { "run": "id-92c4d6b8781d02", "issue": "ending_not_exact", "stage": 6, "profile": "800" }, { "run": "id-d7ac07e53a8a3e", "issue": "ending_not_exact", "stage": 4, "profile": "fixed" }, { "run": "id-bbb0e286d8d5b1", "issue": "ending_not_exact", "stage": 1, "profile": "echo_on" }, { "run": "id-14b5e5d06c59cf", "issue": "ending_not_exact", "stage": 7, "profile": "combined" }, { "run": "id-63250cb7105e3a", "issue": "full_refusal", "stage": 7, "profile": "baseline", "measured_wer": 1.0, "reasons": [] }, { "run": "id-63250cb7105e3a", "issue": "ending_not_exact", "stage": 7, "profile": "baseline" }, { "run": "id-7eb1994f6e47e0", "issue": "ending_not_exact", "stage": 7, "profile": "combined" }, { "run": "id-440164cbae5cdb", "issue": "ending_not_exact", "stage": 2, "profile": "rms24" }, { "run": "id-4c3a210643b27d", "issue": "ending_not_exact", "stage": 2, "profile": "baseline" }, { "run": "id-8d52e37c3c2092", "issue": "ending_not_exact", "stage": 2, "profile": "baseline" }, { "run": "id-ea97207332af91", "issue": "ending_not_exact", "stage": 4, "profile": "causal" }, { "run": "id-203da5c56ea85b", "issue": "ending_not_exact", "stage": 3, "profile": "native" }, { "run": "id-8535c2f1909a20", "issue": "ending_not_exact", "stage": 2, "profile": "rms24" }, { "run": "id-ca3717db46ae7f", "issue": "full_refusal", "stage": 7, "profile": "combined", "measured_wer": 1.0, "reasons": [] }, { "run": "id-ca3717db46ae7f", "issue": "ending_not_exact", "stage": 7, "profile": "combined" }, { "run": "id-f220c5efe62c8d", "issue": "ending_not_exact", "stage": 1, "profile": "echo_off" }, { "run": "id-0c6de530694e87", "issue": "ending_not_exact", "stage": 2, "profile": "rms22" }, { "run": "id-67a16b9eda4f13", "issue": "ending_not_exact", "stage": 4, "profile": "native_agc" }, { "run": "id-2f6ab4c7aff72c", "issue": "ending_not_exact", "stage": 4, "profile": "fixed" }, { "run": "id-5aed0fc76d1f11", "issue": "ending_not_exact", "stage": 3, "profile": "none" }, { "run": "id-5c848a3217b65a", "issue": "ending_not_exact", "stage": 2, "profile": "baseline" }, { "run": "id-25d1d0b04ba4aa", "issue": "full_refusal", "stage": 7, "profile": "combined", "measured_wer": 1.0, "reasons": [ { "owner_ms": 14606.788225471973, "version": 1, "result_type": "error", "session_id": "id-d131543cbb7b00", "stream_id": "id-944742547ffb03", "phrase_id": "id-6b31ea3da4964d", "revision": 17, "text": "", "code": "audio_queue_overflow", "traceparent": "00-7fc0c1426e0fe2aacf80a691faef1548-3e635f0a5b251717-01", "timings": { "version": 1, "asr": { "backend": "streaming_worker", "worker_queue_ms": 1.052219420671463, "worker_decode_ms": 3285.3240575641394, "decode_count": 17, "otel_trace_id": "7fc0c1426e0fe2aacf80a691faef1548", "otel_span_id": "338908670fb0b48c", "last_speech_to_final_ms": 305.8827519416809 } } }, { "owner_ms": 14623.093601316214, "session_id": "id-d131543cbb7b00", "stream_id": "id-944742547ffb03", "phrase_id": "id-6b31ea3da4964d", "version": 1, "result_type": "error", "revision": 1000000, "text": "", "timings": { "version": 1, "asr": { "backend": "streaming_worker" } }, "code": "bridge_failure" } ] }, { "run": "id-25d1d0b04ba4aa", "issue": "ending_not_exact", "stage": 7, "profile": "combined" }, { "run": "id-1f3581ead917b1", "issue": "ending_not_exact", "stage": 4, "profile": "fixed" }, { "run": "id-1b8e38ca639262", "issue": "ending_not_exact", "stage": 4, "profile": "fixed" }, { "run": "id-e7473d4732d462", "issue": "ending_not_exact", "stage": 2, "profile": "baseline" }, { "run": "id-afbcc23f1c99da", "issue": "ending_not_exact", "stage": 6, "profile": "500" }, { "run": "id-d89030a68e7375", "issue": "full_refusal", "stage": 7, "profile": "combined", "measured_wer": 1.0, "reasons": [] }, { "run": "id-d89030a68e7375", "issue": "ending_not_exact", "stage": 7, "profile": "combined" }, { "run": "id-1672f735c30b60", "issue": "ending_not_exact", "stage": 2, "profile": "rms24" }, { "run": "id-7919c039d93f3a", "issue": "full_refusal", "stage": 3, "profile": "rnnoise", "measured_wer": 1.0, "reasons": [ { "owner_ms": 9182.716820389032, "version": 1, "result_type": "error", "session_id": "id-d131543cbb7b00", "stream_id": "id-44e2a47dffe461", "phrase_id": "id-6b5c666f4b7eef", "revision": 4, "text": "", "code": "audio_queue_overflow", "traceparent": "00-528b01e726e3955dcd1a1fa0adf9ba2b-a44141dc0404f20f-01", "timings": { "version": 1, "asr": { "backend": "streaming_worker", "worker_queue_ms": 0.6154943257570267, "worker_decode_ms": 2638.799589127302, "decode_count": 9, "otel_trace_id": "528b01e726e3955dcd1a1fa0adf9ba2b", "otel_span_id": "4aca1d16e48cc102", "last_speech_to_final_ms": 9.250923991203308 } } }, { "owner_ms": 9207.453060895205, "session_id": "id-d131543cbb7b00", "stream_id": "id-44e2a47dffe461", "phrase_id": "id-6b5c666f4b7eef", "version": 1, "result_type": "error", "revision": 1000000, "text": "", "timings": { "version": 1, "asr": { "backend": "streaming_worker" } }, "code": "bridge_failure" } ] }, { "run": "id-7919c039d93f3a", "issue": "ending_not_exact", "stage": 3, "profile": "rnnoise" }, { "run": "id-f12be563f1c189", "issue": "ending_not_exact", "stage": 5, "profile": "250" }, { "run": "id-48b6ecaf628751", "issue": "ending_not_exact", "stage": 2, "profile": "rms22" }, { "run": "id-f3498be1af3cb4", "issue": "ending_not_exact", "stage": 4, "profile": "fixed" }, { "run": "id-27cc6586dbaa03", "issue": "ending_not_exact", "stage": 4, "profile": "native_agc" }, { "run": "id-756b13e1a260f0", "issue": "ending_not_exact", "stage": 6, "profile": "800" }, { "run": "id-72e5b1aa3fca13", "issue": "ending_not_exact", "stage": 2, "profile": "rms24" }, { "run": "id-eaa53364876b0b", "issue": "ending_not_exact", "stage": 3, "profile": "none" }, { "run": "id-b4e705aa97467f", "issue": "ending_not_exact", "stage": 5, "profile": "250" }, { "run": "id-ed64b97d92c7f5", "issue": "full_refusal", "stage": 7, "profile": "baseline", "measured_wer": 1.0, "reasons": [] }, { "run": "id-ed64b97d92c7f5", "issue": "ending_not_exact", "stage": 7, "profile": "baseline" }, { "run": "id-0e6fd06a7dbfcb", "issue": "ending_not_exact", "stage": 1, "profile": "echo_on" }, { "run": "id-9ee5031d8fe789", "issue": "ending_not_exact", "stage": 4, "profile": "native_agc" }, { "run": "id-acdcd320cace66", "issue": "ending_not_exact", "stage": 6, "profile": "no_taps" }, { "run": "id-9fff977ff0c432", "issue": "ending_not_exact", "stage": 7, "profile": "combined" }, { "run": "id-d0170e84defe55", "issue": "ending_not_exact", "stage": 7, "profile": "combined" }, { "run": "id-29cb9027bb8e00", "issue": "ending_not_exact", "stage": 1, "profile": "echo_off" }, { "run": "id-537ddca90a1470", "issue": "ending_not_exact", "stage": 5, "profile": "450" }, { "run": "id-943f27bb34d5b0", "issue": "ending_not_exact", "stage": 4, "profile": "native_agc" }, { "run": "id-a9ceea0a61320f", "issue": "full_refusal", "stage": 7, "profile": "combined", "measured_wer": 1.0, "reasons": [ { "owner_ms": 14185.591388493776, "version": 1, "result_type": "error", "session_id": "id-d131543cbb7b00", "stream_id": "id-d0d0ba1a3a2772", "phrase_id": "id-1469604373221a", "revision": 18, "text": "", "code": "audio_queue_overflow", "traceparent": "00-353415b2b80ccc59047b1d4f0b1ed96f-73f94836b0d1dee7-01", "timings": { "version": 1, "asr": { "backend": "streaming_worker", "worker_queue_ms": 1.17388553917408, "worker_decode_ms": 2958.1872578710318, "decode_count": 17, "otel_trace_id": "353415b2b80ccc59047b1d4f0b1ed96f", "otel_span_id": "1041504cf8fc4635", "last_speech_to_final_ms": 10.383229702711105 } } }, { "owner_ms": 14196.13890722394, "session_id": "id-d131543cbb7b00", "stream_id": "id-d0d0ba1a3a2772", "phrase_id": "id-1469604373221a", "version": 1, "result_type": "error", "revision": 1000000, "text": "", "timings": { "version": 1, "asr": { "backend": "streaming_worker" } }, "code": "bridge_failure" } ] }, { "run": "id-a9ceea0a61320f", "issue": "ending_not_exact", "stage": 7, "profile": "combined" }, { "run": "id-c6ff8ee2d2c863", "issue": "ending_not_exact", "stage": 2, "profile": "rms22" }, { "run": "id-ad9a2140121a34", "issue": "ending_not_exact", "stage": 3, "profile": "rnnoise" }, { "run": "id-8b41055e3ecac9", "issue": "ending_not_exact", "stage": 6, "profile": "500" }, { "run": "id-f1b354d50a5334", "issue": "ending_not_exact", "stage": 2, "profile": "rms24" }, { "run": "id-c592c46435af22", "issue": "ending_not_exact", "stage": 7, "profile": "combined" }, { "run": "id-28102d7794054c", "issue": "ending_not_exact", "stage": 2, "profile": "baseline" }, { "run": "id-f33c769bb7d6ed", "issue": "ending_not_exact", "stage": 7, "profile": "baseline" }, { "run": "id-7121bab4bc5908", "issue": "ending_not_exact", "stage": 7, "profile": "baseline" }, { "run": "id-538a5101ad397e", "issue": "ending_not_exact", "stage": 6, "profile": "taps" }, { "run": "id-635259aa77a73b", "issue": "ending_not_exact", "stage": 3, "profile": "rnnoise" }, { "run": "id-d497e7fb17475d", "issue": "ending_not_exact", "stage": 4, "profile": "causal" }, { "run": "id-acc8bbea9c7075", "issue": "ending_not_exact", "stage": 6, "profile": "1000" }, { "run": "id-17f59fc66de43c", "issue": "ending_not_exact", "stage": 4, "profile": "native_agc" }, { "run": "id-738a69dbab0203", "issue": "ending_not_exact", "stage": 7, "profile": "baseline" }, { "run": "id-221d9ade0f5fac", "issue": "ending_not_exact", "stage": 4, "profile": "fixed" }, { "run": "id-c66d0921911d4e", "issue": "ending_not_exact", "stage": 6, "profile": "500" }, { "run": "id-12e43a3573396e", "issue": "ending_not_exact", "stage": 3, "profile": "none" }, { "run": "id-365d6371d081c8", "issue": "full_refusal", "stage": 1, "profile": "echo_off", "measured_wer": 1.0, "reasons": [ { "owner_ms": 8592.74885430932, "version": 1, "result_type": "error", "session_id": "id-d131543cbb7b00", "stream_id": "id-995c48c42a8efc", "phrase_id": "id-2e8934dcb5d743", "revision": 12, "text": "", "code": "audio_queue_overflow", "traceparent": "00-648f1c4f6a1fcc975c2dd47b656eee01-fb6e574d7d3b3941-01", "timings": { "version": 1, "asr": { "backend": "streaming_worker", "worker_queue_ms": 0.7083844393491745, "worker_decode_ms": 1843.3214202523232, "decode_count": 11, "otel_trace_id": "648f1c4f6a1fcc975c2dd47b656eee01", "otel_span_id": "0b0461fa0220d33e", "last_speech_to_final_ms": 9.76504385471344 } } }, { "owner_ms": 8627.739485353231, "session_id": "id-d131543cbb7b00", "stream_id": "id-995c48c42a8efc", "phrase_id": "id-2e8934dcb5d743", "version": 1, "result_type": "error", "revision": 1000000, "text": "", "timings": { "version": 1, "asr": { "backend": "streaming_worker" } }, "code": "bridge_failure" } ] }, { "run": "id-365d6371d081c8", "issue": "ending_not_exact", "stage": 1, "profile": "echo_off" }, { "run": "id-20b7880f04ba69", "issue": "ending_not_exact", "stage": 2, "profile": "rms24" }, { "run": "id-d7825202e2ecdb", "issue": "ending_not_exact", "stage": 3, "profile": "none" }, { "run": "id-238a3156a8b971", "issue": "full_refusal", "stage": 7, "profile": "baseline", "measured_wer": 1.0, "reasons": [ { "owner_ms": 12331.430464982986, "version": 1, "result_type": "error", "session_id": "id-d131543cbb7b00", "stream_id": "id-017ec2144d4cbf", "phrase_id": "id-e0eda439fc2ee2", "revision": 16, "text": "", "code": "audio_queue_overflow", "traceparent": "00-24eee04d448f34aee0fec662dcf35e47-ef6493e536e093cc-01", "timings": { "version": 1, "asr": { "backend": "streaming_worker", "worker_queue_ms": 0.9030066430568695, "worker_decode_ms": 2489.46382291615, "decode_count": 14, "otel_trace_id": "24eee04d448f34aee0fec662dcf35e47", "otel_span_id": "6374721deae572ed", "last_speech_to_final_ms": 7.304362952709198 } } }, { "owner_ms": 12348.094940185547, "session_id": "id-d131543cbb7b00", "stream_id": "id-017ec2144d4cbf", "phrase_id": "id-e0eda439fc2ee2", "version": 1, "result_type": "error", "revision": 1000000, "text": "", "timings": { "version": 1, "asr": { "backend": "streaming_worker" } }, "code": "bridge_failure" } ] }, { "run": "id-238a3156a8b971", "issue": "ending_not_exact", "stage": 7, "profile": "baseline" }, { "run": "id-c184bb6903c944", "issue": "ending_not_exact", "stage": 5, "profile": "300" }, { "run": "id-0cf157a010d037", "issue": "ending_not_exact", "stage": 7, "profile": "baseline" }, { "run": "id-b95dc277820453", "issue": "ending_not_exact", "stage": 3, "profile": "rnnoise" }, { "run": "id-54a8eb55043dd4", "issue": "ending_not_exact", "stage": 3, "profile": "native" }, { "run": "id-e9e7feab33235b", "issue": "ending_not_exact", "stage": 4, "profile": "causal" }, { "run": "id-e6b2c1ffce32fb", "issue": "ending_not_exact", "stage": 6, "profile": "800" }, { "run": "id-3fdf691fa58c98", "issue": "ending_not_exact", "stage": 5, "profile": "300" }, { "run": "id-f2ac6543d927c5", "issue": "ending_not_exact", "stage": 3, "profile": "rnnoise" }, { "run": "id-b6799400e81406", "issue": "ending_not_exact", "stage": 2, "profile": "rms22" }, { "run": "id-b3d313d0214744", "issue": "ending_not_exact", "stage": 5, "profile": "250" }, { "run": "id-c710bce3cfd25a", "issue": "ending_not_exact", "stage": 2, "profile": "rms22" }, { "run": "id-dc67fed80b3bb7", "issue": "ending_not_exact", "stage": 3, "profile": "none" }, { "run": "id-c8c3118cf8749e", "issue": "ending_not_exact", "stage": 4, "profile": "causal" }, { "run": "id-37809bf1b35d5a", "issue": "ending_not_exact", "stage": 1, "profile": "echo_on" }, { "run": "id-e360b0e4cd3a55", "issue": "ending_not_exact", "stage": 6, "profile": "1000" }, { "run": "id-27cf1f40f60971", "issue": "ending_not_exact", "stage": 7, "profile": "baseline" }, { "run": "id-2ad1bbf8f43b0c", "issue": "full_refusal", "stage": 7, "profile": "baseline", "measured_wer": 1.0, "reasons": [] }, { "run": "id-2ad1bbf8f43b0c", "issue": "ending_not_exact", "stage": 7, "profile": "baseline" }, { "run": "id-7dc497ba00a87c", "issue": "ending_not_exact", "stage": 4, "profile": "causal" }, { "run": "id-1802d5d87721aa", "issue": "ending_not_exact", "stage": 2, "profile": "rms22" }, { "run": "id-8b526a06b37a5c", "issue": "ending_not_exact", "stage": 2, "profile": "baseline" }, { "run": "id-00802f4ab4e974", "issue": "ending_not_exact", "stage": 4, "profile": "causal" }, { "run": "id-39920185a57666", "issue": "full_refusal", "stage": 7, "profile": "combined", "measured_wer": 1.0, "reasons": [] }, { "run": "id-39920185a57666", "issue": "ending_not_exact", "stage": 7, "profile": "combined" }, { "run": "id-fe3c33e3ca120b", "issue": "full_refusal", "stage": 3, "profile": "native", "measured_wer": 1.0, "reasons": [ { "owner_ms": 8589.091375470161, "version": 1, "result_type": "error", "session_id": "id-d131543cbb7b00", "stream_id": "id-dbc21194d29591", "phrase_id": "id-fbf547e1a33536", "revision": 9, "text": "", "code": "audio_queue_overflow", "traceparent": "00-c58d31b1af71c6a322f1def0ca82b0ee-96fd5696d8e097be-01", "timings": { "version": 1, "asr": { "backend": "streaming_worker", "worker_queue_ms": 0.7626134902238846, "worker_decode_ms": 1777.20813639462, "decode_count": 11, "otel_trace_id": "c58d31b1af71c6a322f1def0ca82b0ee", "otel_span_id": "648a1bedbc08bb6b", "last_speech_to_final_ms": 8.914344012737274 } } }, { "owner_ms": 8624.61819127202, "session_id": "id-d131543cbb7b00", "stream_id": "id-dbc21194d29591", "phrase_id": "id-fbf547e1a33536", "version": 1, "result_type": "error", "revision": 1000000, "text": "", "timings": { "version": 1, "asr": { "backend": "streaming_worker" } }, "code": "bridge_failure" } ] }, { "run": "id-fe3c33e3ca120b", "issue": "ending_not_exact", "stage": 3, "profile": "native" }, { "run": "id-4b4a8658d25152", "issue": "ending_not_exact", "stage": 4, "profile": "native_agc" }, { "run": "id-2d26b2b0166439", "issue": "ending_not_exact", "stage": 6, "profile": "1000" }, { "run": "id-fe0a200ea42733", "issue": "ending_not_exact", "stage": 5, "profile": "450" }, { "run": "id-8b85a6962a44de", "issue": "full_refusal", "stage": 3, "profile": "native", "measured_wer": 1.0, "reasons": [] }, { "run": "id-8b85a6962a44de", "issue": "ending_not_exact", "stage": 3, "profile": "native" }, { "run": "id-8844937f78f5ed", "issue": "ending_not_exact", "stage": 3, "profile": "native" } ], "measured_scheduler_boundary_file": "results/isolated-boundaries.json", "limitations": [ "Synthetic acoustic path only; physical room echo not validated", "One screening repetition does not prove a statistical improvement", "Intermediate quality is a separate control re-decode; latency is never reused", "Decode includes runtime work, not pure GPU time", "Stage 6 END/abort/overflow races have separate deterministic scheduler evidence; acoustic live runs do not prove exact overlap" ], "long_repetition_control_rejection": { "control_run": "id-1a1b033240b348", "compression_ratio": 2.5174825174825175, "reasons": [ "compression_ratio_out_of_range" ], "rejected_hypothesis_wer": 0.35135135135135137, "actual_final_wer": 1.0, "scope": "offline full-window control only; rejected hypothesis is not an accepted result" }, "narratives": [ { "stage": 1, "title": "Микрофон и эхо", "paragraphs": [ "Наименьший суммарный online WER в сопоставимых ветках этапа: echo_on, 61.1% при CER 59.6%. Само это число не отменяет остальных критериев отбора.", "Выбран echo_on: WER 61.1%, CER 59.6%; верно распознано 21 из 54 слов, 9 из 12 чисел и 0 из 3 TEST-кодов. Полных отказов: 0; ложных final: 0; дублей: 0. Медиана последняя речь → final: 768 мс.", "Ни один другой кандидат не прошёл все согласованные критерии. Сохраняем baseline; меньший отдельный WER не оправдывает потери окончаний, чисел, разрезание вопроса или отказ.", "Подключение muted: getUserMedia вызван 0 раз. После mute трек: ended; при повторном включении транспорт тот же: True. Поздно полученные треки после отмены: ['ended'].", "Изменение действующего трека через exact applyConstraints отклонено для echoCancellation, noiseSuppression, autoGainControl. Это фактическое ограничение этого Chrome/источника. Профили для акустических сравнений получены новым getUserMedia и проверены через getSettings; успешная смена параметров на действующем треке не заявляется.", "Синтетическое эхо: ослабление RMS после Chrome с AEC относительно без AEC — 72.2 дБ; по 3 / 3 запусков. AGC одинаково отключён. Это внешний показатель стенда, не ERLE и не проверка комнаты.", "Сохранность своей речи при эхе остаётся слабым местом: WER с AEC 61.1%. Выполнение ориентира по ослаблению эха не означает, что полный критерий качества одновременной речи достигнут." ], "selected": "echo_on", "rejected": { "echo_off": [ "new_false_finals", "new_splits_or_finals", "new_full_refusal", "numbers_or_codes_degraded" ] }, "statistics": [ { "stage": 1, "profile": "echo_off", "n": 6, "wer": 0.7407407407407407, "cer": 0.7104377104377104, "false_finals": 6, "full_refusals": 1, "scored_trials": 3, "intentional_aborts": 0, "reference_words": 54, "correct_words": 14, "correct_numbers": 8, "reference_numbers": 12, "correct_codes": 0, "reference_codes": 3, "duplicate_finals": 0, "errors": 0, "terminal_error_runs": 1, "terminal_error_codes": [ "audio_queue_overflow", "bridge_failure" ], "capture_incomplete": 0, "timed_final_trials": 5, "final_p50_ms": 666.1063972860575, "final_min_ms": 644.3984396755695, "final_max_ms": 796.7720404267311, "decode_count": 76 }, { "stage": 1, "profile": "echo_on", "n": 6, "wer": 0.6111111111111112, "cer": 0.5959595959595959, "false_finals": 0, "full_refusals": 0, "scored_trials": 3, "intentional_aborts": 0, "reference_words": 54, "correct_words": 21, "correct_numbers": 9, "reference_numbers": 12, "correct_codes": 0, "reference_codes": 3, "duplicate_finals": 0, "errors": 0, "terminal_error_runs": 0, "terminal_error_codes": [], "capture_incomplete": 0, "timed_final_trials": 3, "final_p50_ms": 768.0222541093826, "final_min_ms": 658.028107136488, "final_max_ms": 837.9631489515305, "decode_count": 44 } ] }, { "stage": 2, "title": "Тихая речь", "paragraphs": [ "Наименьший суммарный online WER в сопоставимых ветках этапа: rms24, 36.0% при CER 21.4%. Само это число не отменяет остальных критериев отбора.", "Выбран baseline: WER 37.4%, CER 23.1%; верно распознано 135 из 203 слов, 4 из 4 чисел и 0 из 1 TEST-кодов. Полных отказов: 0; ложных final: 0; дублей: 0. Медиана последняя речь → final: 676 мс.", "Ни один другой кандидат не прошёл все согласованные критерии. Сохраняем baseline; меньший отдельный WER не оправдывает потери окончаний, чисел, разрезание вопроса или отказ.", "Тихая обычная речь и числа/коды вместе: baseline 45.5%, rms24 45.5%, rms22 41.8%. Оценка включает полные отказы и не заменяет слова цифрами при расчёте WER.", "В тихой фразе baseline сохранил 4 из 4 числовых значений, но 0 из 1 TEST-кодов. Замена записи «тысяча километров» на «1000 км» тоже влияет на строгий WER; поэтому числа, код и буквальный текст показаны раздельно." ], "selected": "baseline", "rejected": { "rms24": [ "quiet_improvement_below_20_percent" ], "rms22": [ "quiet_improvement_below_20_percent" ] }, "statistics": [ { "stage": 2, "profile": "baseline", "n": 6, "wer": 0.37438423645320196, "cer": 0.23071046600458364, "false_finals": 0, "full_refusals": 0, "scored_trials": 6, "intentional_aborts": 0, "reference_words": 203, "correct_words": 135, "correct_numbers": 4, "reference_numbers": 4, "correct_codes": 0, "reference_codes": 1, "duplicate_finals": 0, "errors": 0, "terminal_error_runs": 0, "terminal_error_codes": [], "capture_incomplete": 0, "timed_final_trials": 6, "final_p50_ms": 676.2350238859653, "final_min_ms": 666.0593934357166, "final_max_ms": 703.7342861294746, "decode_count": 162 }, { "stage": 2, "profile": "rms22", "n": 6, "wer": 0.37438423645320196, "cer": 0.226890756302521, "false_finals": 0, "full_refusals": 0, "scored_trials": 6, "intentional_aborts": 0, "reference_words": 203, "correct_words": 136, "correct_numbers": 4, "reference_numbers": 4, "correct_codes": 0, "reference_codes": 1, "duplicate_finals": 0, "errors": 0, "terminal_error_runs": 0, "terminal_error_codes": [], "capture_incomplete": 0, "timed_final_trials": 6, "final_p50_ms": 679.3657364323735, "final_min_ms": 672.1029374748468, "final_max_ms": 792.5739027559757, "decode_count": 164 }, { "stage": 2, "profile": "rms24", "n": 6, "wer": 0.35960591133004927, "cer": 0.21390374331550802, "false_finals": 0, "full_refusals": 0, "scored_trials": 6, "intentional_aborts": 0, "reference_words": 203, "correct_words": 138, "correct_numbers": 4, "reference_numbers": 4, "correct_codes": 0, "reference_codes": 1, "duplicate_finals": 0, "errors": 0, "terminal_error_runs": 0, "terminal_error_codes": [], "capture_incomplete": 0, "timed_final_trials": 6, "final_p50_ms": 671.9748312607408, "final_min_ms": 668.3018133044243, "final_max_ms": 752.3663267493248, "decode_count": 165 } ] }, { "stage": 3, "title": "Шумоподавление", "paragraphs": [ "Наименьший суммарный online WER в сопоставимых ветках этапа: none, 82.0% при CER 68.3%. Само это число не отменяет остальных критериев отбора.", "Выбран native: WER 83.8%, CER 71.6%; верно распознано 41 из 222 слов, 0 из 0 чисел и 0 из 0 TEST-кодов. Полных отказов: 2; ложных final: 0; дублей: 0. Медиана последняя речь → final: 671 мс.", "Ни один другой кандидат не прошёл все согласованные критерии. Сохраняем baseline; меньший отдельный WER не оправдывает потери окончаний, чисел, разрезание вопроса или отказ.", "На чистом синтетическом шуме native снизил RMS выходной записи на 16.5 дБ относительно варианта без NS; ложных final — 0. Это сравнение полного захвата одинаковой длительности, не оценка раздельных компонентов смеси речи и шума. Сильнее подавить чистый шум не означает лучше сохранить слова в смеси.", "На чистом синтетическом шуме rnnoise снизил RMS выходной записи на 31.3 дБ относительно варианта без NS; ложных final — 0. Это сравнение полного захвата одинаковой длительности, не оценка раздельных компонентов смеси речи и шума. Сильнее подавить чистый шум не означает лучше сохранить слова в смеси.", "Вход snr15, none: изменение уровня неречевых участков после Chrome + Worklet -0.1 дБ; речевых участков -14.0 дБ. Это оценка по маске исходника и выравниванию огибающих; падение уровня речи не равно количеству потерянных слов.", "Вход snr15, native: изменение уровня неречевых участков после Chrome + Worklet -6.5 дБ; речевых участков -21.1 дБ. Это оценка по маске исходника и выравниванию огибающих; падение уровня речи не равно количеству потерянных слов.", "Вход snr15, rnnoise: изменение уровня неречевых участков после Chrome + Worklet -3.2 дБ; речевых участков -16.6 дБ. Это оценка по маске исходника и выравниванию огибающих; падение уровня речи не равно количеству потерянных слов." ], "selected": "native", "rejected": { "none": [ "new_splits_or_finals", "clean_codes_or_endings_degraded" ], "rnnoise": [ "new_splits_or_finals", "clean_codes_or_endings_degraded", "no_noise_wer_improvement" ] }, "statistics": [ { "stage": 3, "profile": "native", "n": 8, "wer": 0.8378378378378378, "cer": 0.7155647382920111, "false_finals": 0, "full_refusals": 2, "scored_trials": 6, "intentional_aborts": 0, "reference_words": 222, "correct_words": 41, "correct_numbers": 0, "reference_numbers": 0, "correct_codes": 0, "reference_codes": 0, "duplicate_finals": 0, "errors": 0, "terminal_error_runs": 1, "terminal_error_codes": [ "audio_queue_overflow", "bridge_failure" ], "capture_incomplete": 0, "timed_final_trials": 4, "final_p50_ms": 670.5093719065189, "final_min_ms": 648.7707663327456, "final_max_ms": 713.8287015259266, "decode_count": 117 }, { "stage": 3, "profile": "none", "n": 8, "wer": 0.8198198198198198, "cer": 0.6825068870523416, "false_finals": 0, "full_refusals": 1, "scored_trials": 6, "intentional_aborts": 0, "reference_words": 222, "correct_words": 45, "correct_numbers": 0, "reference_numbers": 0, "correct_codes": 0, "reference_codes": 0, "duplicate_finals": 0, "errors": 0, "terminal_error_runs": 1, "terminal_error_codes": [ "audio_queue_overflow", "bridge_failure" ], "capture_incomplete": 0, "timed_final_trials": 5, "final_p50_ms": 683.1254921853542, "final_min_ms": 647.7418467402458, "final_max_ms": 718.667533248663, "decode_count": 130 }, { "stage": 3, "profile": "rnnoise", "n": 8, "wer": 0.8603603603603603, "cer": 0.7334710743801653, "false_finals": 0, "full_refusals": 1, "scored_trials": 6, "intentional_aborts": 0, "reference_words": 222, "correct_words": 35, "correct_numbers": 0, "reference_numbers": 0, "correct_codes": 0, "reference_codes": 0, "duplicate_finals": 0, "errors": 0, "terminal_error_runs": 1, "terminal_error_codes": [ "audio_queue_overflow", "bridge_failure" ], "capture_incomplete": 0, "timed_final_trials": 5, "final_p50_ms": 668.0963970720768, "final_min_ms": 638.4422443807125, "final_max_ms": 730.3509339690208, "decode_count": 155 } ] }, { "stage": 4, "title": "Подстройка уровня", "paragraphs": [ "Наименьший суммарный online WER в сопоставимых ветках этапа: native_agc, 42.3% при CER 24.8%. Само это число не отменяет остальных критериев отбора.", "Выбран native_agc: WER 42.3%, CER 24.8%; верно распознано 141 из 222 слов, 0 из 0 чисел и 0 из 0 TEST-кодов. Полных отказов: 0; ложных final: 0; дублей: 0. Медиана последняя речь → final: 677 мс.", "Ни один другой кандидат не прошёл все согласованные критерии. Сохраняем baseline; меньший отдельный WER не оправдывает потери окончаний, чисел, разрезание вопроса или отказ.", "Собственный регулятор: максимальное усиление участка, определённого независимой маской как пауза, +17.3 дБ; незавершённых подстроек дольше секунды: 0. Даже быстрая сходимость к собственной цели не доказывает правильность разделения речи и шума.", "Дополнительная проверка заведомо нулевого входа: усиление больше 0,1 дБ найдено в 2 из 6 случаев; максимум +16.4 дБ. Используется выравнивание с шагом 20 мс и защита 100 мс у границ. Энергетическая маска реальных записей может относить тихую речь к паузе, поэтому обе оценки сохранены отдельно." ], "selected": "native_agc", "rejected": { "fixed": [ "new_splits_or_finals", "clean_codes_or_endings_degraded" ], "causal": [ "clean_codes_or_endings_degraded", "source_pause_amplification" ] }, "statistics": [ { "stage": 4, "profile": "causal", "n": 6, "wer": 0.5225225225225225, "cer": 0.3023415977961432, "false_finals": 0, "full_refusals": 0, "scored_trials": 6, "intentional_aborts": 0, "reference_words": 222, "correct_words": 118, "correct_numbers": 0, "reference_numbers": 0, "correct_codes": 0, "reference_codes": 0, "duplicate_finals": 0, "errors": 0, "terminal_error_runs": 0, "terminal_error_codes": [], "capture_incomplete": 0, "timed_final_trials": 6, "final_p50_ms": 696.277148090303, "final_min_ms": 672.9392018169165, "final_max_ms": 753.5087279975414, "decode_count": 177 }, { "stage": 4, "profile": "fixed", "n": 6, "wer": 0.44594594594594594, "cer": 0.2610192837465565, "false_finals": 0, "full_refusals": 0, "scored_trials": 6, "intentional_aborts": 0, "reference_words": 222, "correct_words": 135, "correct_numbers": 0, "reference_numbers": 0, "correct_codes": 0, "reference_codes": 0, "duplicate_finals": 0, "errors": 0, "terminal_error_runs": 0, "terminal_error_codes": [], "capture_incomplete": 0, "timed_final_trials": 6, "final_p50_ms": 690.3958087787032, "final_min_ms": 666.6546799242496, "final_max_ms": 720.8775877952576, "decode_count": 177 }, { "stage": 4, "profile": "native_agc", "n": 6, "wer": 0.42342342342342343, "cer": 0.24793388429752067, "false_finals": 0, "full_refusals": 0, "scored_trials": 6, "intentional_aborts": 0, "reference_words": 222, "correct_words": 141, "correct_numbers": 0, "reference_numbers": 0, "correct_codes": 0, "reference_codes": 0, "duplicate_finals": 0, "errors": 0, "terminal_error_runs": 0, "terminal_error_codes": [], "capture_incomplete": 0, "timed_final_trials": 6, "final_p50_ms": 676.6237840056419, "final_min_ms": 663.9434397220612, "final_max_ms": 686.1139684915543, "decode_count": 178 } ] }, { "stage": 5, "title": "Завершение фразы", "paragraphs": [ "Наименьший суммарный online WER в сопоставимых ветках этапа: 450, 20.4% при CER 17.4%. Само это число не отменяет остальных критериев отбора.", "Выбран 450: WER 20.4%, CER 17.4%; верно распознано 126 из 157 слов, 8 из 8 чисел и 0 из 2 TEST-кодов. Полных отказов: 0; ложных final: 0; дублей: 0. Медиана последняя речь → final: 671 мс.", "Ни один другой кандидат не прошёл все согласованные критерии. Сохраняем baseline; меньший отдельный WER не оправдывает потери окончаний, чисел, разрезание вопроса или отказ.", "Пауза 450 мс: 7 из 8 входов дали больше final, чем заранее задано фраз. Проверка сравнивает и baseline, и абсолютную ожидаемую границу.", "Пауза 300 мс: 8 из 8 входов дали больше final, чем заранее задано фраз. Проверка сравнивает и baseline, и абсолютную ожидаемую границу.", "Пауза 250 мс: 8 из 8 входов дали больше final, чем заранее задано фраз. Проверка сравнивает и baseline, и абсолютную ожидаемую границу.", "Граница интерпретации вставок: 200 мс → приблизительно 420 мс общего тихого участка, 350 мс → приблизительно 570 мс общего тихого участка, 650 мс → приблизительно 870 мс общего тихого участка. Исходная тихая граница около 220 мс сохранена; это оценка по энергии с шагом 20 мс, не точная разметка Silero. Поэтому эти входы не названы чистыми паузами 200/350/650 мс." ], "selected": "450", "rejected": { "300": [ "new_splits_or_finals", "new_phrase_splits", "question_split_against_declared_reference" ], "250": [ "new_splits_or_finals", "new_phrase_splits", "question_split_against_declared_reference" ] }, "statistics": [ { "stage": 5, "profile": "250", "n": 8, "wer": 0.21019108280254778, "cer": 0.16301969365426697, "false_finals": 0, "full_refusals": 0, "scored_trials": 8, "intentional_aborts": 0, "reference_words": 157, "correct_words": 125, "correct_numbers": 8, "reference_numbers": 8, "correct_codes": 0, "reference_codes": 2, "duplicate_finals": 0, "errors": 0, "terminal_error_runs": 0, "terminal_error_codes": [], "capture_incomplete": 0, "timed_final_trials": 8, "final_p50_ms": 440.90111274272203, "final_min_ms": 418.3173570781946, "final_max_ms": 452.54023745656013, "decode_count": 131 }, { "stage": 5, "profile": "300", "n": 8, "wer": 0.21019108280254778, "cer": 0.16301969365426697, "false_finals": 0, "full_refusals": 0, "scored_trials": 8, "intentional_aborts": 0, "reference_words": 157, "correct_words": 125, "correct_numbers": 8, "reference_numbers": 8, "correct_codes": 0, "reference_codes": 2, "duplicate_finals": 0, "errors": 0, "terminal_error_runs": 0, "terminal_error_codes": [], "capture_incomplete": 0, "timed_final_trials": 8, "final_p50_ms": 499.6019033715129, "final_min_ms": 490.2112130075693, "final_max_ms": 516.661811619997, "decode_count": 132 }, { "stage": 5, "profile": "450", "n": 8, "wer": 0.20382165605095542, "cer": 0.1739606126914661, "false_finals": 0, "full_refusals": 0, "scored_trials": 8, "intentional_aborts": 0, "reference_words": 157, "correct_words": 126, "correct_numbers": 8, "reference_numbers": 8, "correct_codes": 0, "reference_codes": 2, "duplicate_finals": 0, "errors": 0, "terminal_error_runs": 0, "terminal_error_codes": [], "capture_incomplete": 0, "timed_final_trials": 8, "final_p50_ms": 671.3884519413114, "final_min_ms": 655.7007934898138, "final_max_ms": 792.8117513656616, "decode_count": 134 } ] }, { "stage": 6, "title": "Whisper и отказы", "paragraphs": [ "Наименьший суммарный online WER в сопоставимых ветках этапа: 500, 25.2% при CER 15.5%. Само это число не отменяет остальных критериев отбора.", "Выбран 500: WER 25.2%, CER 15.5%; верно распознано 108 из 139 слов, 0 из 0 чисел и 0 из 0 TEST-кодов. Полных отказов: 0; ложных final: 0; дублей: 0. Медиана последняя речь → final: 662 мс.", "Ни один другой кандидат не прошёл все согласованные критерии. Сохраняем baseline; меньший отдельный WER не оправдывает потери окончаний, чисел, разрезание вопроса или отказ.", "Длинный вход, interval 500 мс: decode 59; повторно обработано 121.32 с аудио; WER 32.4%; ожидание успешного final 680 мс. Отсутствующий final не считается ускорением.", "Длинный вход, interval 800 мс: decode 42; повторно обработано 75.60 с аудио; WER 32.4%; ожидание успешного final 670 мс. Отсутствующий final не считается ускорением.", "На этом длинном входе 800 мс уменьшил число decode на 28.8% относительно 500 мс. Пригодность общего режима проверяется также на тихой речи и остальных случаях, а не только на этой экономии.", "Длинный вход, interval 1000 мс: decode 39; повторно обработано 70.76 с аудио; WER 33.8%; ожидание успешного final 672 мс. Отсутствующий final не считается ускорением.", "На этом длинном входе 1000 мс уменьшил число decode на 33.9% относительно 500 мс. Пригодность общего режима проверяется также на тихой речи и остальных случаях, а не только на этой экономии.", "Цена серверного съёма, taps: final 709 мс; сумма decode 2405 мс; полный запуск со сбором диагностики 23133 мс; WER 55.6%. По одному запуску на вариант; это наблюдение, а не доказанная величина накладных расходов. Browser capture и диагностический decoder присутствуют в обоих вариантах.", "Цена серверного съёма, no_taps: final 784 мс; сумма decode 2369 мс; полный запуск со сбором диагностики 14594 мс; WER 55.6%. По одному запуску на вариант; это наблюдение, а не доказанная величина накладных расходов. Browser capture и диагностический decoder присутствуют в обоих вариантах.", "Полное контрольное окно длинной повторённой фразы отклонено: compression_ratio_out_of_range; compression_ratio 2.52. У отклонённой гипотезы WER 35.1%, но фактический результат — отказ и WER 100.0%. Этот пример относится к полному контрольному re-decode: online-вариант дал текст. Поэтому переход к полному окну сам по себе не гарантирует улучшения." ], "selected": "500", "rejected": { "800": [ "clean_codes_or_endings_degraded" ], "1000": [ "clean_codes_or_endings_degraded" ] }, "statistics": [ { "stage": 6, "profile": "1000", "n": 7, "wer": 0.26618705035971224, "cer": 0.16289592760180996, "false_finals": 0, "full_refusals": 0, "scored_trials": 4, "intentional_aborts": 1, "reference_words": 139, "correct_words": 108, "correct_numbers": 0, "reference_numbers": 0, "correct_codes": 0, "reference_codes": 0, "duplicate_finals": 0, "errors": 0, "terminal_error_runs": 0, "terminal_error_codes": [], "capture_incomplete": 0, "timed_final_trials": 5, "final_p50_ms": 680.3294643759727, "final_min_ms": 648.8161832094193, "final_max_ms": 770.5769315361977, "decode_count": 79 }, { "stage": 6, "profile": "500", "n": 7, "wer": 0.2517985611510791, "cer": 0.15497737556561086, "false_finals": 0, "full_refusals": 0, "scored_trials": 4, "intentional_aborts": 1, "reference_words": 139, "correct_words": 108, "correct_numbers": 0, "reference_numbers": 0, "correct_codes": 0, "reference_codes": 0, "duplicate_finals": 0, "errors": 0, "terminal_error_runs": 0, "terminal_error_codes": [], "capture_incomplete": 0, "timed_final_trials": 5, "final_p50_ms": 661.8186123669147, "final_min_ms": 645.9328476339579, "final_max_ms": 680.2900172770023, "decode_count": 118 }, { "stage": 6, "profile": "800", "n": 7, "wer": 0.28776978417266186, "cer": 0.15497737556561086, "false_finals": 0, "full_refusals": 0, "scored_trials": 4, "intentional_aborts": 1, "reference_words": 139, "correct_words": 104, "correct_numbers": 0, "reference_numbers": 0, "correct_codes": 0, "reference_codes": 0, "duplicate_finals": 0, "errors": 0, "terminal_error_runs": 0, "terminal_error_codes": [], "capture_incomplete": 0, "timed_final_trials": 5, "final_p50_ms": 666.1147885024548, "final_min_ms": 646.8217819929123, "final_max_ms": 727.2902000695467, "decode_count": 86 }, { "stage": 6, "profile": "no_taps", "n": 1, "wer": 0.5555555555555556, "cer": 0.5252525252525253, "false_finals": 0, "full_refusals": 0, "scored_trials": 1, "intentional_aborts": 0, "reference_words": 18, "correct_words": 8, "correct_numbers": 4, "reference_numbers": 4, "correct_codes": 0, "reference_codes": 1, "duplicate_finals": 0, "errors": 0, "terminal_error_runs": 0, "terminal_error_codes": [], "capture_incomplete": 0, "timed_final_trials": 1, "final_p50_ms": 784.0333878993988, "final_min_ms": 784.0333878993988, "final_max_ms": 784.0333878993988, "decode_count": 15 }, { "stage": 6, "profile": "taps", "n": 1, "wer": 0.5555555555555556, "cer": 0.5252525252525253, "false_finals": 0, "full_refusals": 0, "scored_trials": 1, "intentional_aborts": 0, "reference_words": 18, "correct_words": 8, "correct_numbers": 4, "reference_numbers": 4, "correct_codes": 0, "reference_codes": 1, "duplicate_finals": 0, "errors": 0, "terminal_error_runs": 0, "terminal_error_codes": [], "capture_incomplete": 0, "timed_final_trials": 1, "final_p50_ms": 708.8331952691078, "final_min_ms": 708.8331952691078, "final_max_ms": 708.8331952691078, "decode_count": 15 } ] }, { "stage": 7, "title": "Общая комбинация: отложенный контроль", "paragraphs": [ "Наименьший суммарный online WER в сопоставимых ветках этапа: baseline, 52.6% при CER 41.7%. Само это число не отменяет остальных критериев отбора.", "Выбран baseline: WER 52.6%, CER 41.7%; верно распознано 213 из 441 слов, 12 из 12 чисел и 0 из 3 TEST-кодов. Полных отказов: 4; ложных final: 0; дублей: 0. Медиана последняя речь → final: 660 мс.", "Общая комбинация не прошла отложенный контроль: held_out_wer_regression, lost_ending, new_full_refusal, new_terminal_asr_error.", "После индивидуального отбора фактические параметры общей комбинации совпали с baseline. Штатный AGC в baseline включён значением Chrome по умолчанию, а в комбинации запрошен явно; getSettings подтверждает одинаковое значение. Парные прогоны показывают повторяемость тракта; различие результатов не приписывается изменению настройки, которого не было." ], "selected": "baseline", "rejected": {}, "statistics": [ { "stage": 7, "profile": "baseline", "n": 18, "wer": 0.5260770975056689, "cer": 0.416582406471183, "false_finals": 0, "full_refusals": 4, "scored_trials": 18, "intentional_aborts": 0, "reference_words": 441, "correct_words": 213, "correct_numbers": 12, "reference_numbers": 12, "correct_codes": 0, "reference_codes": 3, "duplicate_finals": 0, "errors": 0, "terminal_error_runs": 1, "terminal_error_codes": [ "audio_queue_overflow", "bridge_failure" ], "capture_incomplete": 0, "timed_final_trials": 14, "final_p50_ms": 659.8912691697478, "final_min_ms": 632.633225992322, "final_max_ms": 705.2855379879475, "decode_count": 244 }, { "stage": 7, "profile": "combined", "n": 18, "wer": 0.6213151927437641, "cer": 0.5426356589147286, "false_finals": 0, "full_refusals": 6, "scored_trials": 18, "intentional_aborts": 0, "reference_words": 441, "correct_words": 167, "correct_numbers": 12, "reference_numbers": 12, "correct_codes": 0, "reference_codes": 3, "duplicate_finals": 0, "errors": 0, "terminal_error_runs": 3, "terminal_error_codes": [ "audio_queue_overflow", "bridge_failure" ], "capture_incomplete": 0, "timed_final_trials": 12, "final_p50_ms": 680.5342324078083, "final_min_ms": 647.6319245994091, "final_max_ms": 754.2067319154739, "decode_count": 240 } ] } ], "weakest_trials": [ { "run": "id-0ce1f63602a83d", "stage": 3, "scenario": 3, "fixture": "s3_3_babble", "profile": "none", "wer": 1.0, "cer": 1.0, "full_refusal": true, "unexpected_finals": 0, "ending_exact": false }, { "run": "id-2edab661426ddd", "stage": 3, "scenario": 1, "fixture": "s3_1_snr5", "profile": "rnnoise", "wer": 1.0, "cer": 0.9504132231404959, "full_refusal": false, "unexpected_finals": 0, "ending_exact": false }, { "run": "id-6abed89bb80ccd", "stage": 7, "scenario": 4, "fixture": "s7_4_transition", "profile": "combined", "wer": 1.0, "cer": 1.0, "full_refusal": true, "unexpected_finals": 0, "ending_exact": false }, { "run": "id-63250cb7105e3a", "stage": 7, "scenario": 3, "fixture": "s7_3_noise", "profile": "baseline", "wer": 1.0, "cer": 1.0, "full_refusal": true, "unexpected_finals": 0, "ending_exact": false }, { "run": "id-ca3717db46ae7f", "stage": 7, "scenario": 3, "fixture": "s7_3_noise", "profile": "combined", "wer": 1.0, "cer": 1.0, "full_refusal": true, "unexpected_finals": 0, "ending_exact": false }, { "run": "id-5aed0fc76d1f11", "stage": 3, "scenario": 1, "fixture": "s3_1_snr5", "profile": "none", "wer": 1.0, "cer": 0.9752066115702479, "full_refusal": false, "unexpected_finals": 0, "ending_exact": false }, { "run": "id-25d1d0b04ba4aa", "stage": 7, "scenario": 1, "fixture": "s7_1_clean", "profile": "combined", "wer": 1.0, "cer": 1.0, "full_refusal": true, "unexpected_finals": 0, "ending_exact": false }, { "run": "id-d89030a68e7375", "stage": 7, "scenario": 3, "fixture": "s7_3_noise", "profile": "combined", "wer": 1.0, "cer": 1.0, "full_refusal": true, "unexpected_finals": 0, "ending_exact": false }, { "run": "id-7919c039d93f3a", "stage": 3, "scenario": 3, "fixture": "s3_3_babble", "profile": "rnnoise", "wer": 1.0, "cer": 1.0, "full_refusal": true, "unexpected_finals": 0, "ending_exact": false }, { "run": "id-ed64b97d92c7f5", "stage": 7, "scenario": 3, "fixture": "s7_3_noise", "profile": "baseline", "wer": 1.0, "cer": 1.0, "full_refusal": true, "unexpected_finals": 0, "ending_exact": false }, { "run": "id-a9ceea0a61320f", "stage": 7, "scenario": 4, "fixture": "s7_4_transition", "profile": "combined", "wer": 1.0, "cer": 1.0, "full_refusal": true, "unexpected_finals": 0, "ending_exact": false }, { "run": "id-365d6371d081c8", "stage": 1, "scenario": 6, "fixture": "s1_6_double_talk", "profile": "echo_off", "wer": 1.0, "cer": 1.0, "full_refusal": true, "unexpected_finals": 0, "ending_exact": false } ], "terminal_errors": [ { "run": "id-0ce1f63602a83d", "stage": 3, "scenario": 3, "profile": "none", "codes": [ "audio_queue_overflow", "bridge_failure" ] }, { "run": "id-6abed89bb80ccd", "stage": 7, "scenario": 4, "profile": "combined", "codes": [ "audio_queue_overflow", "bridge_failure" ] }, { "run": "id-25d1d0b04ba4aa", "stage": 7, "scenario": 1, "profile": "combined", "codes": [ "audio_queue_overflow", "bridge_failure" ] }, { "run": "id-7919c039d93f3a", "stage": 3, "scenario": 3, "profile": "rnnoise", "codes": [ "audio_queue_overflow", "bridge_failure" ] }, { "run": "id-a9ceea0a61320f", "stage": 7, "scenario": 4, "profile": "combined", "codes": [ "audio_queue_overflow", "bridge_failure" ] }, { "run": "id-365d6371d081c8", "stage": 1, "scenario": 6, "profile": "echo_off", "codes": [ "audio_queue_overflow", "bridge_failure" ] }, { "run": "id-238a3156a8b971", "stage": 7, "scenario": 4, "profile": "baseline", "codes": [ "audio_queue_overflow", "bridge_failure" ] }, { "run": "id-fe3c33e3ca120b", "stage": 3, "scenario": 3, "profile": "native", "codes": [ "audio_queue_overflow", "bridge_failure" ] } ], "improvements": [ { "point": "Разделить audio_queue_overflow на превышение длительности окна, накопленного звука и ёмкости очереди; сохранять сработавший предел числом.", "global": "Для длинной неустойчивой гипотезы проверить ограниченное продвижение окна с сохранением слов и приоритетом final; confidence-фильтры не ослаблять.", "status": "В этой серии есть реальные отказы. Последние окна и ожидание очереди сохранены в overflow_evidence; один общий код не подтверждает, какой именно предел сработал. Изменение алгоритма — будущая работа." }, { "point": "Калибровать добавляемый шум и оценку уровня по явно выбранной речевой полосе, отдельно от DC и низких частот исходной записи.", "global": "Измерять спектр до и после браузера и назначать обработке одного владельца; необходимость дополнительного фильтра определять по фактическому PCM.", "status": "В исходниках ниже 80 Гц находится 89,2% / 96,3% энергии, но в двух проверенных baseline-захватах после Chrome осталось 5,1% / 3,6%. Поэтому исходный спектр не доказывает такую же проблему на входе Whisper. Польза дополнительного фильтра для WER не проверялась." }, { "point": "Проверять реальную конфигурацию GUM и освобождение треков, а не только успех applyConstraints.", "global": "Включать поддержку настройки и фактическое значение в диагностику клиента.", "status": "Измерено на функциональных испытаниях; победители не включены в рабочий тракт." }, { "point": "Оценивать сохранность собственной речи при AEC отдельно от сильного подавления чистого эха.", "global": "Перед выводом о качестве реального микрофона провести отдельный согласованный double-talk в комнате с фиксированной геометрией и уровнем громкоговорителя.", "status": "На синтетическом стенде ослабление эха 72,2 дБ, но WER собственной речи при эхе 61,1%. Физическая комната этой кампанией не проверялась." }, { "point": "Для тихой речи выбирать только профиль, который сохраняет числа и окончания; обычным усилением не исправлять уже перегруженный исходник.", "global": "Контроль уровня до записи и отдельная оценка предела микрофона.", "status": "Выбор основан на измерениях; эффект на физическом устройстве требует отдельной проверки." }, { "point": "Не объединять Chrome NS и RNNoise; исключать вариант с неподтверждённым отключением штатного AGC.", "global": "Один владелец DSP и воспроизводимые профили вместо независимых усилителей.", "status": "Сравнение из этой серии; новая production-настройка не применена." }, { "point": "Не сокращать endpointing при новых разрезаниях вопроса; сохранять тихие окончания.", "global": "Возможная адаптация паузы по контексту VAD — отдельная гипотеза, которую эта серия не проверяет.", "status": "Фиксированные паузы проверены; адаптивное завершение не реализовано." }, { "point": "Уменьшать повторные partial decode только при сохранении final и ограниченной очереди.", "global": "Стабильный префикс и приоритет final; измерять повторно обработанные секунды каждого окна.", "status": "Интервалы проверены, алгоритм production не изменён." }, { "point": "При исследовании полного финального окна отдельно проверять настоящие повторы речи против compression_ratio-фильтра.", "global": "Сопоставлять повторяющийся текст с акустическими сегментами; проверять защиту от галлюцинаций и сохранность реальных повторов одним контрольным набором.", "status": "В полном контрольном окне 41,3 с сработал compression_ratio 2,52: гипотеза имела WER 35,1%, фактический отказ — 100%. Online-вариант этого входа дал текст. Фильтры не ослаблялись." }, { "point": "Отдельно проверять буквы в позывном и назначение чисел: сохранённые 123, 1000, 30 и 3500 не компенсируют «Тес» вместо TEST.", "global": "Отдельный контроль критичных терминов и кодов на отложенных голосах; доменные подсказки decoder можно исследовать отдельным кандидатом без автоматической подмены ответа.", "status": "Потеря буквы измерена в тихом baseline. Польза доменной подсказки не проверялась; модель и её параметры в рабочем тракте не менялись." }, { "point": "Диагностический exporter должен проверяться внутри фактического runtime, включая Python 3.10.", "global": "Контрактные проверки образа и явная полнота каждой точки съёма.", "status": "Совместимость exporter исправлена; неполные старые прогоны сохранены и исключены из выбора." } ], "spectral_audit": { "version": 2, "sources": [ { "source": "real1", "mean_dc": -0.02095458178877212, "below_80_hz_energy_fraction": 0.8924879108879078, "below_300_hz_energy_fraction": 0.9754911506315894, "peak_above_full_scale_fraction": 0.0076051323627903695, "representation": "prepared FLOAT WAV; resample overshoot is not proof of ADC clipping" }, { "source": "real2", "mean_dc": -0.03165289220146924, "below_80_hz_energy_fraction": 0.9634101644606503, "below_300_hz_energy_fraction": 0.9902364291893941, "peak_above_full_scale_fraction": 0.012061792549099644, "representation": "prepared FLOAT WAV; resample overshoot is not proof of ADC clipping" }, { "source": "male_ordinary", "mean_dc": 2.3843157320417048e-05, "below_80_hz_energy_fraction": 0.040296576784214194, "below_300_hz_energy_fraction": 0.6164161240730033, "peak_above_full_scale_fraction": 0.0, "representation": "prepared FLOAT WAV; resample overshoot is not proof of ADC clipping" }, { "source": "male_codes", "mean_dc": 5.5489992962203215e-06, "below_80_hz_energy_fraction": 0.0421354758129191, "below_300_hz_energy_fraction": 0.5920684739832416, "peak_above_full_scale_fraction": 0.0, "representation": "prepared FLOAT WAV; resample overshoot is not proof of ADC clipping" }, { "source": "female_ordinary", "mean_dc": -6.512744885655076e-07, "below_80_hz_energy_fraction": 0.0004560596099168953, "below_300_hz_energy_fraction": 0.4351695602466852, "peak_above_full_scale_fraction": 0.0, "representation": "prepared FLOAT WAV; resample overshoot is not proof of ADC clipping" }, { "source": "female_codes", "mean_dc": -4.233428578317901e-06, "below_80_hz_energy_fraction": 0.00044229402128961214, "below_300_hz_energy_fraction": 0.42591604004084754, "peak_above_full_scale_fraction": 0.0, "representation": "prepared FLOAT WAV; resample overshoot is not proof of ADC clipping" } ], "mixtures": [ { "fixture": "s3_1_snr15", "nominal_snr_db": 15, "measured_snr_db": { "full_band": 14.999999590358385, "80_8000_hz": 5.154131745625159, "300_3400_hz": 1.9663424098290792 }, "mask": "same fixed source-energy mask as stimulus preparation" }, { "fixture": "s3_1_snr5", "nominal_snr_db": 5, "measured_snr_db": { "full_band": 4.999999773400649, "80_8000_hz": -4.855277058415234, "300_3400_hz": -8.020374252270955 }, "mask": "same fixed source-energy mask as stimulus preparation" }, { "fixture": "s3_2_keyboard", "nominal_snr_db": 10, "measured_snr_db": { "full_band": 9.999999929036331, "80_8000_hz": 4.806632867632933, "300_3400_hz": 3.0222526942779524 }, "mask": "same fixed source-energy mask as stimulus preparation" }, { "fixture": "s3_3_babble", "nominal_snr_db": 5, "measured_snr_db": { "full_band": 4.999999421744432, "80_8000_hz": -4.767596466204748, "300_3400_hz": -7.238841991667457 }, "mask": "same fixed source-energy mask as stimulus preparation" }, { "fixture": "s7_3_noise", "nominal_snr_db": 5, "measured_snr_db": { "full_band": 4.999999819877683, "80_8000_hz": -10.320303315758185, "300_3400_hz": -13.447970710138586 }, "mask": "same fixed source-energy mask as stimulus preparation" } ], "captured_checkpoints": [ { "run": "id-8e5b62828f8c42", "fixture": "s7_1_clean", "checkpoint": "input", "below_80_hz_energy_fraction": 0.9633908571004018, "mean_dc": -0.01421748596229479, "scope": "one baseline capture; same FFT energy method" }, { "run": "id-8e5b62828f8c42", "fixture": "s7_1_clean", "checkpoint": "chrome", "below_80_hz_energy_fraction": 0.03606462255805962, "mean_dc": 4.835878639011826e-07, "scope": "one baseline capture; same FFT energy method" }, { "run": "id-8e5b62828f8c42", "fixture": "s7_1_clean", "checkpoint": "worklet", "below_80_hz_energy_fraction": 0.03606462255805962, "mean_dc": 4.835878639011826e-07, "scope": "one baseline capture; same FFT energy method" }, { "run": "id-8e5b62828f8c42", "fixture": "s7_1_clean", "checkpoint": "ingress", "below_80_hz_energy_fraction": 0.03639356497642812, "mean_dc": 5.326009308808174e-08, "scope": "one baseline capture; same FFT energy method" }, { "run": "id-8b526a06b37a5c", "fixture": "s2_1_ordinary", "checkpoint": "input", "below_80_hz_energy_fraction": 0.8925068256603423, "mean_dc": -0.019606864697571703, "scope": "one baseline capture; same FFT energy method" }, { "run": "id-8b526a06b37a5c", "fixture": "s2_1_ordinary", "checkpoint": "chrome", "below_80_hz_energy_fraction": 0.050501785771355785, "mean_dc": -1.075138870043618e-05, "scope": "one baseline capture; same FFT energy method" }, { "run": "id-8b526a06b37a5c", "fixture": "s2_1_ordinary", "checkpoint": "worklet", "below_80_hz_energy_fraction": 0.050501785771355785, "mean_dc": -1.075138870043618e-05, "scope": "one baseline capture; same FFT energy method" }, { "run": "id-8b526a06b37a5c", "fixture": "s2_1_ordinary", "checkpoint": "ingress", "below_80_hz_energy_fraction": 0.05063460143637428, "mean_dc": -1.0781288146972656e-05, "scope": "one baseline capture; same FFT energy method" } ], "method": "Known clean component subtracted from the immutable mixture; common headroom gain cancels. Fourth-order Butterworth bands are analysis-only.", "limitations": "Nominal SNR uses full-band source RMS, including low-frequency energy. Band-limited SNR is a separate descriptive metric, not a new stimulus or perceptual quality score.", "audio_modified": false, "asr_runs": 0 }, "pause_audit": { "version": 1, "measurements": [ { "run": "id-d497e7fb17475d", "scenario": 1, "fixture": "s4_1_transition", "available": true, "known_silent_frames": 54, "known_silence_max_applied_gain_db": 0.0, "energy_mask_quiet_max_applied_gain_db": 0.0, "alignment_resolution_ms": 20, "edge_guard_ms": 100 }, { "run": "id-e9e7feab33235b", "scenario": 2, "fixture": "s4_2_transition", "available": true, "known_silent_frames": 36, "known_silence_max_applied_gain_db": 0.0, "energy_mask_quiet_max_applied_gain_db": 17.302003015797204, "alignment_resolution_ms": 20, "edge_guard_ms": 100 }, { "run": "id-00802f4ab4e974", "scenario": 3, "fixture": "s4_3_noise_change", "available": true, "known_silent_frames": 33, "known_silence_max_applied_gain_db": 16.418514998972434, "energy_mask_quiet_max_applied_gain_db": 16.418514998972434, "alignment_resolution_ms": 20, "edge_guard_ms": 100 }, { "run": "id-ea97207332af91", "scenario": 4, "fixture": "s4_4_quiet_ending", "available": true, "known_silent_frames": 35, "known_silence_max_applied_gain_db": 0.0, "energy_mask_quiet_max_applied_gain_db": 13.580323982014646, "alignment_resolution_ms": 20, "edge_guard_ms": 100 }, { "run": "id-c8c3118cf8749e", "scenario": 5, "fixture": "s4_5_pause", "available": true, "known_silent_frames": 56, "known_silence_max_applied_gain_db": 14.815694480436052, "energy_mask_quiet_max_applied_gain_db": 14.815694480436052, "alignment_resolution_ms": 20, "edge_guard_ms": 100 }, { "run": "id-7dc497ba00a87c", "scenario": 6, "fixture": "s4_6_regulators", "available": true, "known_silent_frames": 54, "known_silence_max_applied_gain_db": 0.0, "energy_mask_quiet_max_applied_gain_db": 0.0, "alignment_resolution_ms": 20, "edge_guard_ms": 100 } ], "method": "Known zero-input portions of the fixed WAV, mapped to actual Worklet gains using retained envelope alignment. No new playback or ASR.", "limitation": "Full-band energy masks on the real recordings can label quiet speech as non-speech because low-frequency energy dominates the peak. Mask-based amplification is not proof that an acoustically annotated pause was amplified.", "selection_changed": false }, "overflow_evidence": [ { "run": "id-0ce1f63602a83d", "stage": 3, "fixture": "s3_3_babble", "decode_count": 11, "last_window_audio_seconds": 7.368, "maximum_measured_queue_ms": 0.10898150503635406, "processed_audio_seconds": 42.872, "reprocessed_audio_seconds": 35.504 }, { "run": "id-6abed89bb80ccd", "stage": 7, "fixture": "s7_4_transition", "decode_count": 14, "last_window_audio_seconds": 7.524, "maximum_measured_queue_ms": 0.07399916648864746, "processed_audio_seconds": 61.652, "reprocessed_audio_seconds": 52.108000000000004 }, { "run": "id-25d1d0b04ba4aa", "stage": 7, "fixture": "s7_1_clean", "decode_count": 17, "last_window_audio_seconds": 7.26, "maximum_measured_queue_ms": 0.07693283259868622, "processed_audio_seconds": 65.192, "reprocessed_audio_seconds": 53.312 }, { "run": "id-7919c039d93f3a", "stage": 3, "fixture": "s3_3_babble", "decode_count": 9, "last_window_audio_seconds": 7.144, "maximum_measured_queue_ms": 0.07759593427181244, "processed_audio_seconds": 31.4, "reprocessed_audio_seconds": 24.256 }, { "run": "id-a9ceea0a61320f", "stage": 7, "fixture": "s7_4_transition", "decode_count": 17, "last_window_audio_seconds": 7.656, "maximum_measured_queue_ms": 0.09512156248092651, "processed_audio_seconds": 70.76, "reprocessed_audio_seconds": 59.26400000000001 }, { "run": "id-365d6371d081c8", "stage": 1, "fixture": "s1_6_double_talk", "decode_count": 11, "last_window_audio_seconds": 7.432, "maximum_measured_queue_ms": 0.08290447294712067, "processed_audio_seconds": 43.352000000000004, "reprocessed_audio_seconds": 35.92 }, { "run": "id-238a3156a8b971", "stage": 7, "fixture": "s7_4_transition", "decode_count": 14, "last_window_audio_seconds": 7.48, "maximum_measured_queue_ms": 0.10082684457302094, "processed_audio_seconds": 63.328, "reprocessed_audio_seconds": 53.848 }, { "run": "id-fe3c33e3ca120b", "stage": 3, "fixture": "s3_3_babble", "decode_count": 11, "last_window_audio_seconds": 7.304, "maximum_measured_queue_ms": 0.093020498752594, "processed_audio_seconds": 42.936, "reprocessed_audio_seconds": 35.632 } ], "overflow_limits_from_source": { "revision": "a30a7a3167e40528b482371645c149243299b75b", "file": "py_faster_whisper/src/streaming_asr/worker.py", "maximum_undecoded_age_seconds": 2, "maximum_dirty_samples": 32000, "maximum_online_window_samples": 128000, "sample_rate": 16000, "scheduler_queue_full_uses_same_code": true, "boundary": "verified source branches; current telemetry does not identify the branch for each recorded failure" }, "confidence_summary": { "decode_windows": 2727, "rejected_windows": 204, "reasons": { "compression_ratio_out_of_range": 17, "avg_logprob_below_threshold": 189 }, "boundary": "decode windows, including repeated partial windows; not independent utterances", "filters_changed": false }, "online_vs_full_ingress": [ { "stage": 1, "profile": "echo_off", "paired_trials": 3, "reference_words": 54, "online_wer": 0.7407407407407407, "full_ingress_wer": 0.5740740740740741, "online_full_refusals": 1, "full_ingress_full_refusals": 0, "boundary": "same retained ingress PCM; full-window offline decode versus online VAD/incremental state; not a single-module causal attribution" }, { "stage": 1, "profile": "echo_on", "paired_trials": 3, "reference_words": 54, "online_wer": 0.6111111111111112, "full_ingress_wer": 0.5925925925925926, "online_full_refusals": 0, "full_ingress_full_refusals": 0, "boundary": "same retained ingress PCM; full-window offline decode versus online VAD/incremental state; not a single-module causal attribution" }, { "stage": 2, "profile": "baseline", "paired_trials": 6, "reference_words": 203, "online_wer": 0.37438423645320196, "full_ingress_wer": 0.2955665024630542, "online_full_refusals": 0, "full_ingress_full_refusals": 0, "boundary": "same retained ingress PCM; full-window offline decode versus online VAD/incremental state; not a single-module causal attribution" }, { "stage": 2, "profile": "rms22", "paired_trials": 6, "reference_words": 203, "online_wer": 0.37438423645320196, "full_ingress_wer": 0.28078817733990147, "online_full_refusals": 0, "full_ingress_full_refusals": 0, "boundary": "same retained ingress PCM; full-window offline decode versus online VAD/incremental state; not a single-module causal attribution" }, { "stage": 2, "profile": "rms24", "paired_trials": 6, "reference_words": 203, "online_wer": 0.35960591133004927, "full_ingress_wer": 0.2857142857142857, "online_full_refusals": 0, "full_ingress_full_refusals": 0, "boundary": "same retained ingress PCM; full-window offline decode versus online VAD/incremental state; not a single-module causal attribution" }, { "stage": 3, "profile": "native", "paired_trials": 6, "reference_words": 222, "online_wer": 0.8378378378378378, "full_ingress_wer": 0.7882882882882883, "online_full_refusals": 2, "full_ingress_full_refusals": 0, "boundary": "same retained ingress PCM; full-window offline decode versus online VAD/incremental state; not a single-module causal attribution" }, { "stage": 3, "profile": "none", "paired_trials": 6, "reference_words": 222, "online_wer": 0.8198198198198198, "full_ingress_wer": 0.7522522522522522, "online_full_refusals": 1, "full_ingress_full_refusals": 0, "boundary": "same retained ingress PCM; full-window offline decode versus online VAD/incremental state; not a single-module causal attribution" }, { "stage": 3, "profile": "rnnoise", "paired_trials": 6, "reference_words": 222, "online_wer": 0.8603603603603603, "full_ingress_wer": 0.8828828828828829, "online_full_refusals": 1, "full_ingress_full_refusals": 2, "boundary": "same retained ingress PCM; full-window offline decode versus online VAD/incremental state; not a single-module causal attribution" }, { "stage": 4, "profile": "causal", "paired_trials": 6, "reference_words": 222, "online_wer": 0.5225225225225225, "full_ingress_wer": 0.4144144144144144, "online_full_refusals": 0, "full_ingress_full_refusals": 0, "boundary": "same retained ingress PCM; full-window offline decode versus online VAD/incremental state; not a single-module causal attribution" }, { "stage": 4, "profile": "fixed", "paired_trials": 6, "reference_words": 222, "online_wer": 0.44594594594594594, "full_ingress_wer": 0.40540540540540543, "online_full_refusals": 0, "full_ingress_full_refusals": 0, "boundary": "same retained ingress PCM; full-window offline decode versus online VAD/incremental state; not a single-module causal attribution" }, { "stage": 4, "profile": "native_agc", "paired_trials": 6, "reference_words": 222, "online_wer": 0.42342342342342343, "full_ingress_wer": 0.36936936936936937, "online_full_refusals": 0, "full_ingress_full_refusals": 0, "boundary": "same retained ingress PCM; full-window offline decode versus online VAD/incremental state; not a single-module causal attribution" }, { "stage": 5, "profile": "250", "paired_trials": 8, "reference_words": 157, "online_wer": 0.21019108280254778, "full_ingress_wer": 0.17834394904458598, "online_full_refusals": 0, "full_ingress_full_refusals": 0, "boundary": "same retained ingress PCM; full-window offline decode versus online VAD/incremental state; not a single-module causal attribution" }, { "stage": 5, "profile": "300", "paired_trials": 8, "reference_words": 157, "online_wer": 0.21019108280254778, "full_ingress_wer": 0.17834394904458598, "online_full_refusals": 0, "full_ingress_full_refusals": 0, "boundary": "same retained ingress PCM; full-window offline decode versus online VAD/incremental state; not a single-module causal attribution" }, { "stage": 5, "profile": "450", "paired_trials": 8, "reference_words": 157, "online_wer": 0.20382165605095542, "full_ingress_wer": 0.19745222929936307, "online_full_refusals": 0, "full_ingress_full_refusals": 0, "boundary": "same retained ingress PCM; full-window offline decode versus online VAD/incremental state; not a single-module causal attribution" }, { "stage": 6, "profile": "1000", "paired_trials": 4, "reference_words": 139, "online_wer": 0.26618705035971224, "full_ingress_wer": 0.60431654676259, "online_full_refusals": 0, "full_ingress_full_refusals": 1, "boundary": "same retained ingress PCM; full-window offline decode versus online VAD/incremental state; not a single-module causal attribution" }, { "stage": 6, "profile": "500", "paired_trials": 4, "reference_words": 139, "online_wer": 0.2517985611510791, "full_ingress_wer": 0.6115107913669064, "online_full_refusals": 0, "full_ingress_full_refusals": 1, "boundary": "same retained ingress PCM; full-window offline decode versus online VAD/incremental state; not a single-module causal attribution" }, { "stage": 6, "profile": "800", "paired_trials": 4, "reference_words": 139, "online_wer": 0.28776978417266186, "full_ingress_wer": 0.6115107913669064, "online_full_refusals": 0, "full_ingress_full_refusals": 1, "boundary": "same retained ingress PCM; full-window offline decode versus online VAD/incremental state; not a single-module causal attribution" }, { "stage": 6, "profile": "taps", "paired_trials": 1, "reference_words": 18, "online_wer": 0.5555555555555556, "full_ingress_wer": 0.5555555555555556, "online_full_refusals": 0, "full_ingress_full_refusals": 0, "boundary": "same retained ingress PCM; full-window offline decode versus online VAD/incremental state; not a single-module causal attribution" }, { "stage": 7, "profile": "baseline", "paired_trials": 18, "reference_words": 441, "online_wer": 0.5260770975056689, "full_ingress_wer": 0.5238095238095238, "online_full_refusals": 4, "full_ingress_full_refusals": 0, "boundary": "same retained ingress PCM; full-window offline decode versus online VAD/incremental state; not a single-module causal attribution" }, { "stage": 7, "profile": "combined", "paired_trials": 18, "reference_words": 441, "online_wer": 0.6213151927437641, "full_ingress_wer": 0.54421768707483, "online_full_refusals": 6, "full_ingress_full_refusals": 0, "boundary": "same retained ingress PCM; full-window offline decode versus online VAD/incremental state; not a single-module causal attribution" } ] }