| """Whisper candidates (MIT) via transformers: ~/venv/bin/python asr_whisper.py <hf_id> <tag>""" |
| import json, sys, time, torch, soundfile as sf |
| from transformers import pipeline |
| mid, tag = sys.argv[1], sys.argv[2]; D = "/scratch/duplex/asr_eval" |
| WL = {"eng": "en", "yor": "yo", "hau": "ha", "pcm": "en", "ibo": None, "cs": None} |
| asr = pipeline("automatic-speech-recognition", model=mid, torch_dtype=torch.float16, device="cuda") |
| R = [json.loads(l) for l in open(f"{D}/list.jsonl")] |
| out = open(f"{D}/hyp_{tag}.jsonl", "w") |
| for i, r in enumerate(R): |
| w = sf.read(r["wav"], dtype="float32")[0] |
| gk = {"task": "transcribe"} |
| if WL[r["lang"]]: gk["language"] = WL[r["lang"]] |
| torch.cuda.synchronize(); t = time.perf_counter() |
| h = asr({"raw": w, "sampling_rate": 16000}, generate_kwargs=gk)["text"].strip() |
| torch.cuda.synchronize(); dt = time.perf_counter() - t |
| out.write(json.dumps(dict(id=r["id"], hyp=h, lat=round(dt, 4), dur=r["dur"]), ensure_ascii=False) + "\n"); out.flush() |
| print("WHISPER_DONE", tag) |
|
|
