codec/scripts/eval_asr_from_filelist.py

import argparse
import json
from pathlib import Path

import nemo.collections.asr as nemo_asr
import torch
import yaml
from torchaudio import load
from torchaudio.functional import resample
from tqdm import tqdm


def load_config(config_path):
    with open(config_path, "r") as f:
        return yaml.safe_load(f)


def transcribe(audio: torch.Tensor, asr_model) -> str:
    audio = audio.cpu().numpy(force=True)
    with torch.inference_mode():
        return asr_model.transcribe([audio[0]])[0].text


def main(args):
    config = load_config(args.config)

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    # Load ASR model
    asr_model = nemo_asr.models.ASRModel.from_pretrained(
        model_name=config.get("asr_model", "nvidia/parakeet-tdt-0.6b-v2")
    )

    # Read file list
    with open(config["file_list"], "r") as f:
        wav_files = [line.strip() for line in f if line.strip()]

    results = []

    for wav_path in tqdm(wav_files, desc="Transcribing"):
        wav, sr = load(wav_path)
        wav = resample(wav, orig_freq=sr, new_freq=16000).to(device)

        transcript = transcribe(wav, asr_model)
        results.append({"file": wav_path, "transcript": transcript})

    # Save output
    out_path = Path(config.get("output_jsonl", "asr_transcripts.jsonl"))
    with out_path.open("w") as f:
        for entry in results:
            f.write(json.dumps(entry, ensure_ascii=False) + "\n")

    print(f"\nSaved {len(results)} transcripts to {out_path}")


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--config", type=str, required=True, help="Path to YAML config")
    args = parser.parse_args()
    main(args)