Add Kyutai causal-streaming predictions and benchmark results

PACKAGE_AUDIT.json

@@ -6,29 +6,29 @@
   "parquet_rows": 625,
   "per_condition": {
     "far_field": {
-      "parquet_sha256": "7f99ef391945e8cf4629cf86f2c06384e9ebd134456f9fe44abb67df0f82f0c8",
+      "parquet_sha256": "cb301cd6306601da8e0c1f28abc42452af024baeeeea9dc679b4d49720c0ab4e",
       "rows": 125
     },
     "far_field_noise": {
-      "parquet_sha256": "778b57f3e5d3e9f5e346318d9c28ef4252f41ddf338a26bb6694a925535116a9",
+      "parquet_sha256": "c9056a2911f9f6fd74ae3151a1cf380432afe6047efe99f75aacae1a7da51d8f",
       "rows": 125
     },
     "noise": {
-      "parquet_sha256": "719a3f4780414248947315894e6b2d8d695b8a2ae17c0ac97afd7b5126f56545",
+      "parquet_sha256": "2026c0ee5c4ab5217bee4479b3ccba807a6aa40a903fdb3e3ac74b1e0628bba5",
       "rows": 125
     },
     "obstructed_noise": {
-      "parquet_sha256": "e6a75899d70b619dc3b10eafe8fe537cce91e55fb99412d2bc33e03130e037db",
+      "parquet_sha256": "b6a00148759d8ca60cd6b2a405bb152c89ac59561f2e13f2d2fa5bfa2f01adff",
       "rows": 125
     },
     "recording_noise": {
-      "parquet_sha256": "75c3beecd0dd844241e12ac7afa7d11f33ccf1242000abde584b279109d78eca",
+      "parquet_sha256": "61477bf329b4bd0ab22da86310e2e2fbb76b67d3599b41790175e8e9bccee1ff",
       "rows": 125
     }
   },
-  "predictions": 1250,
+  "predictions": 1875,
   "status": "PASS",
-  "systems": 2,
+  "systems": 3,
   "unique_audio_hashes": 625,
   "unique_scene_ids": 625
 }

README.md

@@ -95,21 +95,27 @@ documented in `results/benchmark_summary.json`.
 |---|---:|
 | A5S | **21.02%** |
 | ElevenLabs Scribe v2 Realtime | 23.52% |
+| Kyutai STT 1B EN/FR | 57.41% |
 
-| Condition | N | A5S | ElevenLabs |
-|---|---:|---:|---:|
-| `far_field` | 125 | 6.06% | 5.51% |
-| `far_field_noise` | 125 | 32.61% | 33.74% |
-| `noise` | 125 | 22.14% | 28.68% |
-| `obstructed_noise` | 125 | 24.97% | 29.53% |
-| `recording_noise` | 125 | 9.53% | 10.14% |
+| Condition | N | A5S | ElevenLabs | Kyutai |
+|---|---:|---:|---:|---:|
+| `far_field` | 125 | 6.06% | 5.51% | 20.70% |
+| `far_field_noise` | 125 | 32.61% | 33.74% | 86.24% |
+| `noise` | 125 | 22.14% | 28.68% | 54.46% |
+| `obstructed_noise` | 125 | 24.97% | 29.53% | 66.22% |
+| `recording_noise` | 125 | 9.53% | 10.14% | 37.54% |
 
 A5S used cache-aware greedy streaming at exactly 560 ms, BF16, and batch size 1.
 ElevenLabs used Scribe v2 Realtime with forced English, one independent WebSocket
 session per conversation, mono PCM16 at 16 kHz, and 100 ms chunks paced in
-wall-clock real time.
-
-Each row contains raw predictions and utterance-level S/D/I counts for both
+wall-clock real time. Kyutai used the official causal Mimi-to-LMGen transcript
+path with 80 ms audio frames, a configured 0.5 s text delay, BF16, temperature 0,
+batch size 1, and fresh streaming state for every conversation. The pinned
+PyTorch checkpoint exposes no semantic-VAD extra heads, so no transcript-gating
+VAD was active. Kyutai returned a blank transcript for 24.48% of conversations;
+these count as deletions in WER.
+
+Each row contains raw predictions and utterance-level S/D/I counts for all three
 systems. Aggregate results live under `results/`.
 
 ## Loading
@@ -121,6 +127,7 @@ dataset = load_dataset("AirCaps/mega-asr-conversational-overlap")
 example = dataset["noise"][0]
 print(example["text"])
 print(example["a5s_prediction"])
+print(example["kyutai_prediction"])
 ```
 
 Recent versions of `datasets` may require `torchcodec` for decoded audio. Use
@@ -134,7 +141,7 @@ are needed.
 - Pinned upstream revision: `a8a35d3319737190d6fd3d39157b258eaab35980`
 - Construction date: 2026-08-24
 - Source usage: 1,250 / 1,250, with zero within-condition repetitions
-- Final audit: 625 decoded WAVs, 1,250 matched predictions, zero missing or corrupt rows
+- Final audit: 625 decoded WAVs, 1,875 matched predictions, zero missing or corrupt rows
 
 Because this is derived from a training corpus rather than a speaker-disjoint
 held-out benchmark, it should be treated as a fixed diagnostic suite. Do not

SHA256SUMS

@@ -1,10 +1,11 @@
-136f72b36299e875d7d5421e7457f89585936b0e200b5ad152ef732ea06136b3  PACKAGE_AUDIT.json
-63134a10b93af58149078252108639e7c45788e6aae2d2863db00b52232a4598  README.md
-7f99ef391945e8cf4629cf86f2c06384e9ebd134456f9fe44abb67df0f82f0c8  data/far_field.parquet
-778b57f3e5d3e9f5e346318d9c28ef4252f41ddf338a26bb6694a925535116a9  data/far_field_noise.parquet
-719a3f4780414248947315894e6b2d8d695b8a2ae17c0ac97afd7b5126f56545  data/noise.parquet
-e6a75899d70b619dc3b10eafe8fe537cce91e55fb99412d2bc33e03130e037db  data/obstructed_noise.parquet
-75c3beecd0dd844241e12ac7afa7d11f33ccf1242000abde584b279109d78eca  data/recording_noise.parquet
+5f880ca4e67d419fb27bb81796cd1b92d25e214d8183578d186e5e83c4c46f4e  PACKAGE_AUDIT.json
+c67a9972de8cf3741bb3abbaec737dbab362e276de5122a3ad541232659f9a9e  README.md
+cb301cd6306601da8e0c1f28abc42452af024baeeeea9dc679b4d49720c0ab4e  data/far_field.parquet
+c9056a2911f9f6fd74ae3151a1cf380432afe6047efe99f75aacae1a7da51d8f  data/far_field_noise.parquet
+2026c0ee5c4ab5217bee4479b3ccba807a6aa40a903fdb3e3ac74b1e0628bba5  data/noise.parquet
+b6a00148759d8ca60cd6b2a405bb152c89ac59561f2e13f2d2fa5bfa2f01adff  data/obstructed_noise.parquet
+61477bf329b4bd0ab22da86310e2e2fbb76b67d3599b41790175e8e9bccee1ff  data/recording_noise.parquet
 e3f369de3b0d598e074c6f33252f8810f1f76239eddd1fd0a25e966394762c3d  provenance/CONSTRUCTION.json
-774f5bc5aa98a8f3df517c6c3139bbc8f33bf2722efd026d48dea751ea27ecce  results/benchmark_summary.csv
-1015251d9be6acf7b1d1d411e14b99809948c6b159a93fc2d4e56647b0182a7d  results/benchmark_summary.json
+e91f7c08a70d50bb9fd97eaa3d34045a2d67f0c7deb0d8e6c766e2198543eff2  provenance/KYUTAI_EVALUATION.json
+7df9934ac4a8e5bc454e1516e9ee723e3216c800535da8d0fee2ab6f87872c6b  results/benchmark_summary.csv
+94fb5b48c7b4777fa3cd33d91c8e80b9d6ddca48cdd7c46d57adfbee9b998a3b  results/benchmark_summary.json

data/far_field.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:7f99ef391945e8cf4629cf86f2c06384e9ebd134456f9fe44abb67df0f82f0c8
-size 80133930
+oid sha256:cb301cd6306601da8e0c1f28abc42452af024baeeeea9dc679b4d49720c0ab4e
+size 80169593

data/far_field_noise.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:778b57f3e5d3e9f5e346318d9c28ef4252f41ddf338a26bb6694a925535116a9
-size 120236686
+oid sha256:c9056a2911f9f6fd74ae3151a1cf380432afe6047efe99f75aacae1a7da51d8f
+size 120253549

data/noise.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:719a3f4780414248947315894e6b2d8d695b8a2ae17c0ac97afd7b5126f56545
-size 88962956
+oid sha256:2026c0ee5c4ab5217bee4479b3ccba807a6aa40a903fdb3e3ac74b1e0628bba5
+size 88993025

data/obstructed_noise.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e6a75899d70b619dc3b10eafe8fe537cce91e55fb99412d2bc33e03130e037db
-size 78890811
+oid sha256:b6a00148759d8ca60cd6b2a405bb152c89ac59561f2e13f2d2fa5bfa2f01adff
+size 78914917

data/recording_noise.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:75c3beecd0dd844241e12ac7afa7d11f33ccf1242000abde584b279109d78eca
-size 74693280
+oid sha256:61477bf329b4bd0ab22da86310e2e2fbb76b67d3599b41790175e8e9bccee1ff
+size 74724006

provenance/KYUTAI_EVALUATION.json

@@ -0,0 +1,73 @@
+{
+  "evaluation": {
+    "audio_hash_mismatches": 0,
+    "blank_predictions": 153,
+    "bucket_mismatches": 0,
+    "conditions": {
+      "far_field": 125,
+      "far_field_noise": 125,
+      "noise": 125,
+      "obstructed_noise": 125,
+      "recording_noise": 125
+    },
+    "duration_mismatches": 0,
+    "examples": 625,
+    "reference_mismatches": 0,
+    "runtime_failures": 0
+  },
+  "inference": {
+    "batch_size": 1,
+    "causal": true,
+    "configured_text_delay_seconds": 0.5,
+    "implementation": "official moshi.run_inference Mimi.encode -> LMGen.step loop",
+    "model_frame_rate_hz": 12.5,
+    "model_input_sample_rate_hz": 24000,
+    "right_padding": "official run_inference rule: audio_delay_seconds + 1.0 seconds",
+    "semantic_vad": false,
+    "semantic_vad_note": "The pinned PyTorch checkpoint exposes zero LM extra heads. The official --vad path only observes extra-head EOT probabilities and does not gate text; this run uses LMGen.step, i.e. the VAD-disabled transcript path.",
+    "semantic_vad_observer_requested": false,
+    "state_isolation": "Mimi and LMGen streaming states reset before every utterance",
+    "temperature_audio": 0.0,
+    "temperature_text": 0.0
+  },
+  "model_files": {
+    "config.json": {
+      "bytes": 1261,
+      "sha256": "5e6aee8007000d844152932184fb7c7112d73cb9a4be562a313b6fa2b89a68f9"
+    },
+    "mimi-pytorch-e351c8d8@125.safetensors": {
+      "bytes": 384644900,
+      "sha256": "09b782f0629851a271227fb9d36db65c041790365f11bbe5d3d59369cf863f50"
+    },
+    "model.safetensors": {
+      "bytes": 1978423520,
+      "sha256": "8f6e244d44baf63c6fa3587d25a4e8d3627ecbafe177cf34a736e697bb725116"
+    },
+    "tokenizer_en_fr_audio_8000.model": {
+      "bytes": 120378,
+      "sha256": "cd87dd5d17169151782ac700280ec057e5d658a9afbe238a048ea5ff318cce69"
+    }
+  },
+  "model_repo": "kyutai/stt-1b-en_fr",
+  "model_revision": "1c34c6b4f7e9299bb61985f145052ff131005dde",
+  "runtime_versions": {
+    "device": "cuda",
+    "dtype": "torch.bfloat16",
+    "jiwer": "4.0.0",
+    "moshi": "0.2.13",
+    "python": "3.11.10 (main, Sep  7 2024, 18:35:41) [GCC 11.4.0]",
+    "sphn": "0.2.1",
+    "torch": "2.9.1+cu128"
+  },
+  "scoring": {
+    "aggregation": "corpus-level WER",
+    "normalizer": "Unicode NFKC; lowercase; normalize curly/backtick apostrophes then remove apostrophes; replace other punctuation with spaces; collapse whitespace"
+  },
+  "source_result_hashes": {
+    "FINAL_AUDIT.json": "5c27c00d6091d21c74eaf75f9081adc9a4717337231bc2ee40648b2da069b148",
+    "MATCHED_PUBLIC_PREDICTIONS.jsonl": "495b18b5201d0e3c56fae2aa568f6aa6a090191f7c645a73d95287d047e5d24e",
+    "PUBLIC_COMPARISON.json": "0c462026bf9304c47557deec9eb3df7fe94f910936c4cd3a2b0f525785ab7f81"
+  },
+  "status": "PASS",
+  "system": "Kyutai STT 1B EN/FR"
+}

results/benchmark_summary.csv

@@ -25,3 +25,16 @@ elevenlabs,ElevenLabs Scribe v2 Realtime,gain_difference_bucket,0_to_lt2_db,257,
 elevenlabs,ElevenLabs Scribe v2 Realtime,gain_difference_bucket,2_to_lt4_db,222,1.3242257291666673,11301,8664,1623,1014,223,2860,25.307494911954695,0.9009009009009009,12.162162162162161
 elevenlabs,ElevenLabs Scribe v2 Realtime,gain_difference_bucket,4_to_lt6_db,113,0.7040244618055553,6221,4938,811,472,134,1417,22.77768847452178,0.0,7.964601769911504
 elevenlabs,ElevenLabs Scribe v2 Realtime,gain_difference_bucket,6_to_8_db,33,0.21326578124999998,1893,1511,208,174,16,398,21.024828314844164,0.0,6.0606060606060606
+kyutai,Kyutai STT 1B EN/FR,overall,overall,625,3.835099461805552,32928,14970,4668,13290,945,18903,57.407069970845484,24.48,0.64
+kyutai,Kyutai STT 1B EN/FR,condition,far_field,125,0.6938356250000002,5608,4656,738,214,209,1161,20.70256776034237,2.4,1.6
+kyutai,Kyutai STT 1B EN/FR,condition,far_field_noise,125,1.0412593402777777,9675,1425,683,7567,94,8344,86.24289405684755,61.6,0.0
+kyutai,Kyutai STT 1B EN/FR,condition,noise,125,0.770268420138889,7135,3409,1060,2666,160,3886,54.46391030133147,18.4,0.8
+kyutai,Kyutai STT 1B EN/FR,condition,obstructed_noise,125,0.6830526388888892,5462,2098,1266,2098,253,3617,66.22116440864153,30.4,0.0
+kyutai,Kyutai STT 1B EN/FR,condition,recording_noise,125,0.6466834375,5048,3382,921,745,229,1895,37.53961965134707,9.6,0.8
+kyutai,Kyutai STT 1B EN/FR,overlap_bin,overlap_50_to_100ms,210,1.262150989583333,10771,4686,1551,4534,319,6404,59.45594652307121,21.904761904761905,0.9523809523809523
+kyutai,Kyutai STT 1B EN/FR,overlap_bin,overlap_gt100_to_200ms,210,1.27174515625,10813,5022,1520,4271,311,6102,56.43207250531767,25.714285714285715,0.0
+kyutai,Kyutai STT 1B EN/FR,overlap_bin,overlap_gt200_to_300ms,205,1.3012033159722218,11344,5262,1597,4485,315,6397,56.391043723554304,25.853658536585368,0.975609756097561
+kyutai,Kyutai STT 1B EN/FR,gain_difference_bucket,0_to_lt2_db,257,1.593583489583332,13513,6216,1869,5428,359,7656,56.65655294901206,24.124513618677042,0.0
+kyutai,Kyutai STT 1B EN/FR,gain_difference_bucket,2_to_lt4_db,222,1.3242257291666673,11301,5183,1800,4318,410,6528,57.76479957525883,23.873873873873872,0.9009009009009009
+kyutai,Kyutai STT 1B EN/FR,gain_difference_bucket,4_to_lt6_db,113,0.7040244618055553,6221,2522,755,2944,122,3821,61.42099340941971,29.20353982300885,1.7699115044247788
+kyutai,Kyutai STT 1B EN/FR,gain_difference_bucket,6_to_8_db,33,0.21326578124999998,1893,1049,244,600,54,898,47.43792921288959,15.151515151515152,0.0

results/benchmark_summary.json

@@ -5,7 +5,8 @@
   "normalizer": "Unicode NFKC; lowercase; normalize curly/backtick apostrophes then remove apostrophes; replace other punctuation with spaces; collapse whitespace",
   "protocols": {
     "a5s": "cache-aware greedy streaming, 560 ms chunks/context, BF16, batch 1",
-    "elevenlabs": "Scribe v2 Realtime; forced English; independent session per scene; mono PCM16 16 kHz; 100 ms chunks paced in wall-clock real time"
+    "elevenlabs": "Scribe v2 Realtime; forced English; independent session per scene; mono PCM16 16 kHz; 100 ms chunks paced in wall-clock real time",
+    "kyutai": "Official causal Mimi -> LMGen path; 80 ms frames; configured 0.5 s text delay; BF16; temperature 0; batch 1; fresh state per conversation; no semantic-VAD gating"
   },
   "status": "PASS",
   "systems": {
@@ -364,6 +365,237 @@
         "wer_pct": 23.517978620019438
       },
       "system": "ElevenLabs Scribe v2 Realtime"
+    },
+    "kyutai": {
+      "by_condition": {
+        "far_field": {
+          "audio_hours": 0.6938356250000002,
+          "blank_pct": 2.4,
+          "compute_rtf": 0.10300705742163713,
+          "deletions": 214,
+          "errors": 1161,
+          "exact_pct": 1.6,
+          "examples": 125,
+          "hits": 4656,
+          "inference_seconds": 257.291877835989,
+          "insertions": 209,
+          "reference_words": 5608,
+          "substitutions": 738,
+          "throughput_x_realtime": 9.708072679978772,
+          "wer_pct": 20.70256776034237
+        },
+        "far_field_noise": {
+          "audio_hours": 1.0412593402777777,
+          "blank_pct": 61.6,
+          "compute_rtf": 0.09871145284689975,
+          "deletions": 7567,
+          "errors": 8344,
+          "exact_pct": 0.0,
+          "examples": 125,
+          "hits": 1425,
+          "inference_seconds": 370.02320016920567,
+          "insertions": 94,
+          "reference_words": 9675,
+          "substitutions": 683,
+          "throughput_x_realtime": 10.130536742793034,
+          "wer_pct": 86.24289405684755
+        },
+        "noise": {
+          "audio_hours": 0.770268420138889,
+          "blank_pct": 18.4,
+          "compute_rtf": 0.10051511161189904,
+          "deletions": 2666,
+          "errors": 3886,
+          "exact_pct": 0.8,
+          "examples": 125,
+          "hits": 3409,
+          "inference_seconds": 278.7250183969736,
+          "insertions": 160,
+          "reference_words": 7135,
+          "substitutions": 1060,
+          "throughput_x_realtime": 9.948752818990249,
+          "wer_pct": 54.46391030133147
+        },
+        "obstructed_noise": {
+          "audio_hours": 0.6830526388888892,
+          "blank_pct": 30.4,
+          "compute_rtf": 0.10129849841796311,
+          "deletions": 2098,
+          "errors": 3617,
+          "exact_pct": 0.0,
+          "examples": 125,
+          "hits": 2098,
+          "inference_seconds": 249.09194397553802,
+          "insertions": 253,
+          "reference_words": 5462,
+          "substitutions": 1266,
+          "throughput_x_realtime": 9.871814643035927,
+          "wer_pct": 66.22116440864153
+        },
+        "recording_noise": {
+          "audio_hours": 0.6466834375,
+          "blank_pct": 9.6,
+          "compute_rtf": 0.10166343494678864,
+          "deletions": 745,
+          "errors": 1895,
+          "exact_pct": 0.8,
+          "examples": 125,
+          "hits": 3382,
+          "inference_seconds": 236.67861448600888,
+          "insertions": 229,
+          "reference_words": 5048,
+          "substitutions": 921,
+          "throughput_x_realtime": 9.836378246745323,
+          "wer_pct": 37.53961965134707
+        }
+      },
+      "by_gain_difference_bucket": {
+        "0_to_lt2_db": {
+          "audio_hours": 1.593583489583332,
+          "blank_pct": 24.124513618677042,
+          "compute_rtf": 0.1011517427615704,
+          "deletions": 5428,
+          "errors": 7656,
+          "exact_pct": 0.0,
+          "examples": 257,
+          "hits": 6216,
+          "inference_seconds": 580.2974899467081,
+          "insertions": 359,
+          "reference_words": 13513,
+          "substitutions": 1869,
+          "throughput_x_realtime": 9.886137131192566,
+          "wer_pct": 56.65655294901206
+        },
+        "2_to_lt4_db": {
+          "audio_hours": 1.3242257291666673,
+          "blank_pct": 23.873873873873872,
+          "compute_rtf": 0.10070255621896959,
+          "deletions": 4318,
+          "errors": 6528,
+          "exact_pct": 0.9009009009009009,
+          "examples": 222,
+          "hits": 5183,
+          "inference_seconds": 480.0704973768443,
+          "insertions": 410,
+          "reference_words": 11301,
+          "substitutions": 1800,
+          "throughput_x_realtime": 9.93023451982272,
+          "wer_pct": 57.76479957525883
+        },
+        "4_to_lt6_db": {
+          "audio_hours": 0.7040244618055553,
+          "blank_pct": 29.20353982300885,
+          "compute_rtf": 0.10044629289196041,
+          "deletions": 2944,
+          "errors": 3821,
+          "exact_pct": 1.7699115044247788,
+          "examples": 113,
+          "hits": 2522,
+          "inference_seconds": 254.57993025705218,
+          "insertions": 122,
+          "reference_words": 6221,
+          "substitutions": 755,
+          "throughput_x_realtime": 9.955569003184573,
+          "wer_pct": 61.42099340941971
+        },
+        "6_to_8_db": {
+          "audio_hours": 0.21326578124999998,
+          "blank_pct": 15.151515151515152,
+          "compute_rtf": 0.10011339011480361,
+          "deletions": 600,
+          "errors": 898,
+          "exact_pct": 0.0,
+          "examples": 33,
+          "hits": 1049,
+          "inference_seconds": 76.86273728311062,
+          "insertions": 54,
+          "reference_words": 1893,
+          "substitutions": 244,
+          "throughput_x_realtime": 9.988673831275358,
+          "wer_pct": 47.43792921288959
+        }
+      },
+      "by_overlap_bin": {
+        "overlap_50_to_100ms": {
+          "audio_hours": 1.262150989583333,
+          "blank_pct": 21.904761904761905,
+          "compute_rtf": 0.10064477627635869,
+          "deletions": 4534,
+          "errors": 6404,
+          "exact_pct": 0.9523809523809523,
+          "examples": 210,
+          "hits": 4686,
+          "inference_seconds": 457.3040543049574,
+          "insertions": 319,
+          "reference_words": 10771,
+          "substitutions": 1551,
+          "throughput_x_realtime": 9.935935445413659,
+          "wer_pct": 59.45594652307121
+        },
+        "overlap_gt100_to_200ms": {
+          "audio_hours": 1.27174515625,
+          "blank_pct": 25.714285714285715,
+          "compute_rtf": 0.10153194209642508,
+          "deletions": 4271,
+          "errors": 6102,
+          "exact_pct": 0.0,
+          "examples": 210,
+          "hits": 5022,
+          "inference_seconds": 464.84192003682256,
+          "insertions": 311,
+          "reference_words": 10813,
+          "substitutions": 1520,
+          "throughput_x_realtime": 9.849117227072227,
+          "wer_pct": 56.43207250531767
+        },
+        "overlap_gt200_to_300ms": {
+          "audio_hours": 1.3012033159722218,
+          "blank_pct": 25.853658536585368,
+          "compute_rtf": 0.10026289485637789,
+          "deletions": 4485,
+          "errors": 6397,
+          "exact_pct": 0.975609756097561,
+          "examples": 205,
+          "hits": 5262,
+          "inference_seconds": 469.6646805219352,
+          "insertions": 315,
+          "reference_words": 11344,
+          "substitutions": 1597,
+          "throughput_x_realtime": 9.973779446847763,
+          "wer_pct": 56.391043723554304
+        }
+      },
+      "model_repo": "kyutai/stt-1b-en_fr",
+      "model_revision": "1c34c6b4f7e9299bb61985f145052ff131005dde",
+      "overall": {
+        "audio_hours": 3.835099461805552,
+        "blank_pct": 24.48,
+        "compute_rtf": 0.10080939872507501,
+        "deletions": 13290,
+        "errors": 18903,
+        "exact_pct": 0.64,
+        "examples": 625,
+        "hits": 14970,
+        "inference_seconds": 1391.8106548637152,
+        "insertions": 945,
+        "reference_words": 32928,
+        "substitutions": 4668,
+        "throughput_x_realtime": 9.919709993779213,
+        "wer_pct": 57.407069970845484
+      },
+      "streaming": {
+        "algorithmic_text_delay_seconds": 0.5,
+        "audio_delivery": "one complete 80 ms Mimi frame at a time",
+        "batch_size": 1,
+        "genuinely_causal": true,
+        "model_config_causal": true,
+        "precision": "BF16",
+        "semantic_vad": false,
+        "semantic_vad_available_in_checkpoint": false,
+        "semantic_vad_effect_on_text": "none; official observer path does not gate transcript tokens",
+        "semantic_vad_observer_requested": false
+      },
+      "system": "Kyutai STT 1B EN/FR"
     }
   }
 }