vh
a6c1d3c454
A/B of the live STT seat (fv-ml1 GPU 0, sherpa-onnx int8 v3) against nvidia/parakeet-unified-en-0.6b, measured on GPU 3 with the seat's own image, k2-fsa's published unified int8 export, fp32/fp16 exports made with k2-fsa's recipe, v2 int8, and NeMo 3.0.0 (fp32, bf16 autocast, bf16 weights). - Seat int8 graph runs on one CPU thread (cpu/wall 1.00, GPU 2-9%). - unified-en under NeMo: -121/-234/-530 ms vs the seat at 1-3/3-8/8-20 s (paired, n=120/bin; floor <=6 ms; +50 ms positive control reads +52-54). - unified-en WER lower in every runtime: -0.7 pp clean, -1.5 pp other, -3.2 to -4.4 pp AMI (paired CIs exclude 0). - Seat defects found: hard 400 s input ceiling (HTTP 500), truncation after a quiet 1.5 s pause, and severe long-window dropouts (int8 v3 only). - B-bf16w needs +0.8 to +1.5 GB over the seat's 1,690 MiB on GPU 0. Raw requests, hypotheses, manifests and the full harness under services/parakeet-ab-2026-09-30/. No deploy; live seat untouched apart from 240 light test requests.
11 lines
2.4 KiB
Plaintext
11 lines
2.4 KiB
Plaintext
{"arm": "ab-a1", "file": "wilde", "windows": 5, "failed": 0, "wer": 11.54, "ref_words": 3580, "dropouts": 10, "dropout_words": 320, "clean_dropouts": 10, "clean_dropout_words": 320, "insertion_runs": 0, "insertion_words": 0, "e2e_ms_median": 17629.211}
|
|
{"arm": "ab-a1", "file": "scotus", "windows": 6, "failed": 0, "wer": 7.03, "ref_words": 5772, "dropouts": 8, "dropout_words": 120, "clean_dropouts": 5, "clean_dropout_words": 74, "insertion_runs": 0, "insertion_words": 0, "e2e_ms_median": 17804.285}
|
|
{"arm": "ab-af32", "file": "wilde", "windows": 5, "failed": 0, "wer": 2.68, "ref_words": 3580, "dropouts": 1, "dropout_words": 16, "clean_dropouts": 1, "clean_dropout_words": 16, "insertion_runs": 0, "insertion_words": 0, "e2e_ms_median": 2451.737}
|
|
{"arm": "ab-af32", "file": "scotus", "windows": 6, "failed": 0, "wer": 3.9, "ref_words": 5772, "dropouts": 3, "dropout_words": 36, "clean_dropouts": 1, "clean_dropout_words": 11, "insertion_runs": 0, "insertion_words": 0, "e2e_ms_median": 1702.969}
|
|
{"arm": "ab-b16w", "file": "wilde", "windows": 5, "failed": 0, "wer": 2.09, "ref_words": 3580, "dropouts": 0, "dropout_words": 0, "clean_dropouts": 0, "clean_dropout_words": 0, "insertion_runs": 0, "insertion_words": 0, "e2e_ms_median": 636.527}
|
|
{"arm": "ab-b16w", "file": "scotus", "windows": 6, "failed": 0, "wer": 4.83, "ref_words": 5772, "dropouts": 3, "dropout_words": 37, "clean_dropouts": 0, "clean_dropout_words": 0, "insertion_runs": 0, "insertion_words": 0, "e2e_ms_median": 633.802}
|
|
{"arm": "ab-c", "file": "wilde", "windows": 5, "failed": 0, "wer": 2.37, "ref_words": 3580, "dropouts": 0, "dropout_words": 0, "clean_dropouts": 0, "clean_dropout_words": 0, "insertion_runs": 0, "insertion_words": 0, "e2e_ms_median": 18178.125}
|
|
{"arm": "ab-c", "file": "scotus", "windows": 6, "failed": 0, "wer": 5.08, "ref_words": 5772, "dropouts": 4, "dropout_words": 48, "clean_dropouts": 0, "clean_dropout_words": 0, "insertion_runs": 0, "insertion_words": 0, "e2e_ms_median": 18120.522}
|
|
{"arm": "ab-d", "file": "wilde", "windows": 5, "failed": 0, "wer": 29.05, "ref_words": 3580, "dropouts": 3, "dropout_words": 1009, "clean_dropouts": 3, "clean_dropout_words": 1009, "insertion_runs": 0, "insertion_words": 0, "e2e_ms_median": 17343.929}
|
|
{"arm": "ab-d", "file": "scotus", "windows": 6, "failed": 0, "wer": 4.95, "ref_words": 5772, "dropouts": 4, "dropout_words": 43, "clean_dropouts": 1, "clean_dropout_words": 12, "insertion_runs": 0, "insertion_words": 0, "e2e_ms_median": 17934.451}
|