eval_model.py

#!/usr/bin/env python3
"""
eval_model.py

Evaluate a .nemo ASR model on a manifest (batch + streaming WER).

Usage:
  python eval_model.py --model /path/to/best_model.nemo --manifest /path/to/test_manifest.json
  python eval_model.py --model /path/to/best_model.nemo --manifest /path/to/test_manifest.json --gpu 0
  python eval_model.py --model /path/to/best_model.nemo --manifest /path/to/test_manifest.json --no-streaming

Manifest format (one JSON object per line):
  {"audio_filepath": "/abs/path/utt.wav", "text": "reference transcript"}
  Only `audio_filepath` and `text` are read; other keys (e.g. `duration`) are ignored.

Note:
    WER here uses Whisper's BasicMultilingualTextNormalizer from the
    Open ASR Leaderboard repo, matching the paper pipeline.

Requirements:
  pip install nemo_toolkit[asr] soundfile numpy
"""

import argparse
import json
import os
import sys

import numpy as np
import torch


# Use Whisper's BasicMultilingualTextNormalizer for consistency with paper runs.
try:
    from normalizer import BasicMultilingualTextNormalizer
    _ml_normalizer = BasicMultilingualTextNormalizer()
except ImportError:
    print(
        "ERROR: could not import BasicMultilingualTextNormalizer. "
        "This script requires Whisper's BasicMultilingualTextNormalizer, "
        "shipped in the Open ASR Leaderboard repo. "
        "Clone https://github.com/huggingface/open_asr_leaderboard and add it "
        "to PYTHONPATH (or set OPEN_ASR_LB_ROOT).",
        file=sys.stderr,
    )
    raise SystemExit(1)


def normalize_text(text):
    return _ml_normalizer(text)


def simple_wer(ref_words, hyp_words):
    n, m = len(ref_words), len(hyp_words)
    dp = [[0] * (m + 1) for _ in range(n + 1)]
    for i in range(n + 1): dp[i][0] = i
    for j in range(m + 1): dp[0][j] = j
    for i in range(1, n + 1):
        for j in range(1, m + 1):
            dp[i][j] = dp[i-1][j-1] if ref_words[i-1] == hyp_words[j-1] \
                else 1 + min(dp[i-1][j], dp[i][j-1], dp[i-1][j-1])
    return dp[n][m]


@torch.no_grad()
def evaluate_batch(model, manifest_path, device):
    import soundfile as sf
    model.eval()

    samples = []
    with open(manifest_path) as f:
        for line in f:
            samples.append(json.loads(line))

    total_edits, total_words = 0, 0
    errors = 0
    batch_size = 16
    examples = []

    for start in range(0, len(samples), batch_size):
        batch_samples = samples[start:start + batch_size]
        try:
            audios = []
            for s in batch_samples:
                audio, sr = sf.read(s["audio_filepath"], dtype="float32")
                if len(audio.shape) > 1:
                    audio = audio.mean(axis=1)
                audios.append(torch.FloatTensor(audio))

            audio_lens = torch.LongTensor([len(a) for a in audios])
            max_len = audio_lens.max().item()
            padded = torch.zeros(len(audios), max_len)
            for i, a in enumerate(audios):
                padded[i, :len(a)] = a

            padded = padded.to(device)
            audio_lens = audio_lens.to(device)

            mel, mel_len = model.preprocessor(input_signal=padded, length=audio_lens)
            enc, enc_len = model.encoder(audio_signal=mel, length=mel_len)

            best_hyps = model.decoding.rnnt_decoder_predictions_tensor(enc, enc_len)
            if isinstance(best_hyps, tuple):
                best_hyps = best_hyps[0]

            for s, hyp in zip(batch_samples, best_hyps):
                if hasattr(hyp, 'text') and hyp.text:
                    pred = hyp.text
                elif hasattr(hyp, 'y_sequence'):
                    tids = hyp.y_sequence.tolist() if torch.is_tensor(hyp.y_sequence) else list(hyp.y_sequence)
                    pred = model.tokenizer.ids_to_text(tids) if tids else ""
                else:
                    pred = str(hyp)

                ref_n = normalize_text(s["text"])
                pred_n = normalize_text(pred)
                ref_words = ref_n.split()
                pred_words = pred_n.split()
                if ref_words:
                    total_edits += simple_wer(ref_words, pred_words)
                    total_words += len(ref_words)

                if len(examples) < 10:
                    examples.append((s["text"][:60], pred[:60]))

        except Exception as e:
            errors += 1
            if errors <= 3:
                print(f"  [batch eval error] {type(e).__name__}: {e}")

    wer_score = total_edits / max(total_words, 1) * 100

    print(f"\n  {'Reference':<60} | {'Prediction':<60}")
    print(f"  {'-'*60} | {'-'*60}")
    for ref, pred in examples:
        print(f"  {ref:<60} | {pred:<60}")
    if errors:
        print(f"  ({errors} batch eval errors)")

    return wer_score, total_edits, total_words


@torch.no_grad()
def evaluate_streaming(model, manifest_path, device):
    import soundfile as sf
    from nemo.collections.asr.parts.utils.streaming_utils import CacheAwareStreamingAudioBuffer

    model.eval()

    right_context = 13
    chunk_frames = 1 + right_context
    model.encoder.setup_streaming_params(
        chunk_size=chunk_frames,
        shift_size=chunk_frames,
        left_chunks=70 // max(chunk_frames, 1),
    )

    samples = []
    with open(manifest_path) as f:
        for line in f:
            samples.append(json.loads(line))

    total_edits, total_words = 0, 0
    examples = []
    errors = 0

    for s in samples:
        try:
            audio, sr = sf.read(s["audio_filepath"], dtype="float32")
            if len(audio.shape) > 1:
                audio = audio.mean(axis=1)

            buffer = CacheAwareStreamingAudioBuffer(model=model)
            buffer.append_audio(audio)

            cache_last_channel, cache_last_time, cache_last_channel_len = \
                model.encoder.get_initial_cache_state(batch_size=1, dtype=torch.float32, device=device)
            previous_hypotheses = None
            pred = ""

            for chunk_audio, chunk_len in buffer:
                if chunk_audio is None:
                    break
                result = model.conformer_stream_step(
                    processed_signal=chunk_audio,
                    processed_signal_length=chunk_len,
                    cache_last_channel=cache_last_channel,
                    cache_last_time=cache_last_time,
                    cache_last_channel_len=cache_last_channel_len,
                    previous_hypotheses=previous_hypotheses,
                    return_transcription=True,
                )
                if isinstance(result, tuple) and len(result) >= 6:
                    cache_last_channel = result[2]
                    cache_last_time = result[3]
                    cache_last_channel_len = result[4]
                    previous_hypotheses = result[5]
                    if result[5] and len(result[5]) > 0:
                        hyp = result[5][0]
                        new_text = ""
                        if hasattr(hyp, 'text') and hyp.text:
                            new_text = hyp.text
                        elif hasattr(hyp, 'y_sequence'):
                            tids = hyp.y_sequence.tolist() if torch.is_tensor(hyp.y_sequence) else list(hyp.y_sequence)
                            if tids:
                                new_text = model.tokenizer.ids_to_text(tids)
                        if new_text and len(new_text) > len(pred):
                            pred = new_text

            ref_n = normalize_text(s["text"])
            pred_n = normalize_text(pred)
            ref_words = ref_n.split()
            pred_words = pred_n.split()

            if ref_words:
                total_edits += simple_wer(ref_words, pred_words)
                total_words += len(ref_words)

            if len(examples) < 10:
                examples.append((s["text"][:60], pred[:60]))

        except Exception as e:
            errors += 1
            if errors <= 3:
                print(f"  [streaming eval error] {type(e).__name__}: {e}")

    wer_score = total_edits / max(total_words, 1) * 100

    print(f"\n  {'Reference':<60} | {'Prediction':<60}")
    print(f"  {'-'*60} | {'-'*60}")
    for ref, pred in examples:
        print(f"  {ref:<60} | {pred:<60}")
    if errors:
        print(f"  ({errors} samples failed)")

    return wer_score, total_edits, total_words


def main():
    parser = argparse.ArgumentParser(description="Evaluate NeMo ASR model")
    parser.add_argument("--model", type=str, required=True, help="Path to .nemo model")
    parser.add_argument("--manifest", type=str, required=True,
                        help="Path to evaluation manifest (JSONL)")
    parser.add_argument("--gpu", type=int, default=0, help="GPU index")
    parser.add_argument("--no-streaming", action="store_true", help="Skip streaming eval")
    args = parser.parse_args()

    device = torch.device(f"cuda:{args.gpu}")
    torch.cuda.set_device(args.gpu)

    import nemo.collections.asr as nemo_asr
    from nemo.core.classes.common import typecheck
    typecheck.set_typecheck_enabled(False)

    print(f"\n{'='*65}")
    print(f"  Model:    {args.model}")
    print(f"  Manifest: {args.manifest}")
    print(f"  GPU:      {args.gpu}")
    print(f"{'='*65}")

    print(f"\n  Loading model...")
    model = nemo_asr.models.ASRModel.restore_from(args.model, map_location=device)
    model = model.to(device)
    model.eval()

    from omegaconf import open_dict
    with open_dict(model.cfg):
        model.cfg.decoding.greedy.use_cuda_graph_decoder = False
    model.change_decoding_strategy(model.cfg.decoding)

    print(f"  Vocab: {model.tokenizer.vocab_size} tokens")
    print(f"  Params: {sum(p.numel() for p in model.parameters()) / 1e6:.1f}M")

    # Count samples
    with open(args.manifest) as f:
        n_samples = sum(1 for _ in f)
    print(f"  Samples: {n_samples}")

    # Batch eval
    print(f"\n{'='*65}")
    print(f"  Batch Evaluation")
    print(f"{'='*65}")
    batch_wer, batch_edits, batch_words = evaluate_batch(model, args.manifest, device)
    print(f"\n  Batch WER: {batch_wer:.2f}% ({batch_edits}/{batch_words})")

    # Streaming eval
    if not args.no_streaming:
        print(f"\n{'='*65}")
        print(f"  Streaming Evaluation")
        print(f"{'='*65}")
        stream_wer, stream_edits, stream_words = evaluate_streaming(model, args.manifest, device)
        print(f"\n  Streaming WER: {stream_wer:.2f}% ({stream_edits}/{stream_words})")

    # Summary
    print(f"\n{'='*65}")
    print(f"  Summary")
    print(f"{'='*65}")
    print(f"  Model:          {os.path.basename(args.model)}")
    print(f"  Manifest:       {os.path.basename(args.manifest)}")
    print(f"  Batch WER:      {batch_wer:.2f}%")
    if not args.no_streaming:
        print(f"  Streaming WER:  {stream_wer:.2f}%")
    print(f"{'='*65}")


if __name__ == "__main__":
    main()