VAP-OTO-CoreML / benchmark.json
aufklarer's picture
Publish validated Float32 export and matched TurnBench pilot
8e62170 verified
Raw History Blame Contribute Delete
5.49 kB
{
"status": "passed",
"conversations": [
"20",
"21"
],
"positive_events": 148,
"negative_spans": 94,
"plan": {
"protocol": "turn-export-pilot-20261030-v1",
"conversations": [
"20",
"21"
],
"dataset": "mundo-ai/turn-benchmark-dev",
"dataset_revision": "8fa18a24be51528a45397b35cbcaecd84202062b",
"scorer_revision": "38a6f874322430cb3ca71d8a52aa1e636e88bad8",
"gold_sha256": "89f7a4cec6b22a8fdc7703ad1b93a603ee752cef7d096a72cd50ff5e3f919b38",
"audio_manifest_sha256": "dada2f249539f67c43879e07298d6211aeabfe0a307814ca3c4cad0f7f57b2f9",
"p99": {
"input": "independent mono channels",
"step_samples": 2560,
"gate": "saved Silero VAD onset 0.5 / offset 0.35",
"minimum_silence_seconds": 0.3,
"maximum_silence_seconds": 2.0,
"threshold": 0.5,
"readout": "latest output available at a completed 160 ms step",
"end_of_stream_flush": false
},
"vap": {
"input": "stereo channel order speaker_1, speaker_2",
"window_samples": 320000,
"readout_samples": 8000,
"readout": "last frame only, timestamped at input window end",
"short_context": "leading zeros",
"eot_threshold": 0.9161,
"refractory_seconds": 2.0,
"onnx_threads": 2
},
"threshold_tuning_on_pilot": false,
"graphs": {
"p99_coreml": "7b7d37a333e098671ddc428d5f5055acc6fbdf1d92383198d815692031201b67",
"vap_coreml": "3d856c039e8aff382b71280546ffeaee5e886ddd8804fe1d75be378968e1fd29",
"vap_onnx": "8dc413028f9181115ef46eb3260cd2be01aa61cf123711843071e45db5b7133a"
}
},
"metrics": {
"p99-coreml": {
"recall": 0.9054054054054054,
"fpr": 0.1595744680851064,
"tp": 134,
"fn": 14,
"fp": 15,
"tn": 79,
"p10_ms": 300.99999999999784,
"p50_ms": 1506.0000000000002,
"p90_ms": 2058.9999999999914,
"runtime": {
"requests": 18958,
"model_compute_seconds": 65.37459301610943,
"compute_rtf_per_conversation": 0.043102675508419036,
"request_ms_p50": 3.0113745015114546,
"request_ms_p95": 5.207720579346642
},
"predictions_sha256": "3ea95d48d1710faf1de6e47ac05652009348da0b065de68e5890f4e1321769e3"
},
"p99-reference": {
"recall": 0.9054054054054054,
"fpr": 0.1595744680851064,
"tp": 134,
"fn": 14,
"fp": 15,
"tn": 79,
"p10_ms": 300.99999999999784,
"p50_ms": 1506.0000000000002,
"p90_ms": 2058.9999999999914,
"runtime": {
"requests": 18958,
"model_compute_seconds": 46.38250518415589,
"compute_rtf_per_conversation": 0.03058084154692321,
"request_ms_p50": 1.9732710206881166,
"request_ms_p95": 4.183964408002792
},
"predictions_sha256": "3ea95d48d1710faf1de6e47ac05652009348da0b065de68e5890f4e1321769e3"
},
"vap-coreml": {
"recall": 0.8378378378378378,
"fpr": 0.02127659574468085,
"tp": 124,
"fn": 24,
"fp": 2,
"tn": 92,
"p10_ms": 388.9000000000072,
"p50_ms": 945.4999999999956,
"p90_ms": 2032.1999999999894,
"runtime": {
"requests": 3032,
"model_compute_seconds": 371.5537947830744,
"compute_rtf_per_conversation": 0.24497227304359984,
"request_ms_p50": 113.04749950068071,
"request_ms_p95": 216.1974496615583
},
"predictions_sha256": "7374a1c9409eb97f6d9ced9ee376ae4654931fe5bd4aaae8e11b67c684780e18"
},
"vap-onnx": {
"recall": 0.8378378378378378,
"fpr": 0.02127659574468085,
"tp": 124,
"fn": 24,
"fp": 2,
"tn": 92,
"p10_ms": 388.9000000000072,
"p50_ms": 945.4999999999956,
"p90_ms": 2032.1999999999894,
"runtime": {
"requests": 3032,
"model_compute_seconds": 1158.3884125971235,
"compute_rtf_per_conversation": 0.7637468557331274,
"request_ms_p50": 377.92920850915834,
"request_ms_p95": 416.1279473511968
},
"predictions_sha256": "7374a1c9409eb97f6d9ced9ee376ae4654931fe5bd4aaae8e11b67c684780e18"
},
"vap-reference": {
"recall": 0.8378378378378378,
"fpr": 0.02127659574468085,
"tp": 124,
"fn": 24,
"fp": 2,
"tn": 92,
"p10_ms": 388.9000000000072,
"p50_ms": 945.4999999999956,
"p90_ms": 2032.1999999999894,
"runtime": {
"requests": 3032,
"model_compute_seconds": 1423.4309210996144,
"compute_rtf_per_conversation": 0.9384942723190351,
"request_ms_p50": 409.97106197755784,
"request_ms_p95": 703.7348371639378
},
"predictions_sha256": "7374a1c9409eb97f6d9ced9ee376ae4654931fe5bd4aaae8e11b67c684780e18"
}
},
"maximum_probability_errors": {
"p99_coreml": 5.270540714263916e-05,
"vap_coreml": 1.1324882507324219e-05,
"vap_onnx": 9.59634780883789e-06
},
"event_predictions_identical_to_reference": {
"p99-coreml": true,
"vap-coreml": true,
"vap-onnx": true
},
"latency_scope": "Offline audio-timeline replay. Request computation is measured separately and excluded from event timestamps.",
"hardware": "Apple M5 Pro, 48 GiB; CPU; ONNX Runtime 2 threads",
"validation_versions": {
"torch": "2.11.0",
"torchaudio": "2.11.0",
"numpy": "2.2.6",
"coremltools": "9.0",
"onnxruntime": "1.31.0"
},
"saved_Silero_trace_sha256": "6cdf85f35d4e5d0384af85f725c0a896fe1ffd7c4d9b250883560bb9dc5ff216"
}