| """omniASR candidates on the eval set: transcripts + single-clip latency + CTC streaming (prefix re-decode) sim. |
| ~/venv_omni/bin/python asr_omni.py <card> -> /scratch/duplex/asr_eval/hyp_<card>.jsonl""" |
| import json, sys, time, numpy as np, soundfile as sf, torch |
| from omnilingual_asr.models.inference.pipeline import ASRInferencePipeline |
| LC = {"ibo": "ibo_Latn", "yor": "yor_Latn", "hau": "hau_Latn", "pcm": "pcm_Latn", "eng": "eng_Latn"} |
| card = sys.argv[1]; D = "/scratch/duplex/asr_eval" |
| R = [json.loads(l) for l in open(f"{D}/list.jsonl")] |
| p = ASRInferencePipeline(model_card=card) |
| ctc = "CTC" in card |
| def tr(w, lang): |
| kw = {} if ctc or lang is None else dict(lang=[LC[lang]]) |
| return p.transcribe([{"waveform": w, "sample_rate": 16000}], batch_size=1, **kw)[0] |
| w0 = sf.read(R[0]["wav"], dtype="float32")[0] |
| for _ in range(3): tr(w0, "eng") |
| out = open(f"{D}/hyp_{card}.jsonl", "w") |
| for r in R: |
| w = sf.read(r["wav"], dtype="float32")[0] |
| lang = r["lang"] if r["lang"] != "cs" else None |
| torch.cuda.synchronize(); t = time.perf_counter(); h = tr(w, lang); torch.cuda.synchronize(); dt = time.perf_counter() - t |
| rec = dict(id=r["id"], hyp=h, lat=round(dt, 4), dur=r["dur"]) |
| if ctc and r["dur"] > 4 and int(r["id"][-3:]) % 4 == 0: |
| parts = [] |
| for e in np.arange(0.64, r["dur"], 0.64): |
| t = time.perf_counter(); hp = tr(w[: int(e * 16000)], None); parts.append((round(float(e), 2), hp, round(time.perf_counter() - t, 4))) |
| rec["partials"] = parts |
| out.write(json.dumps(rec, ensure_ascii=False) + "\n"); out.flush() |
| print("OMNI_DONE", card) |
|
|
