🐛 Bug
ASR transcription failed with special token validation error during batch dataset inference, while most samples run normally with stable RTF performance.
To Reproduce
code: https://www.kaggle.com/code/liuweiq/funasr-recog-script?scriptVersionId=333049663
Code sample
hotword = load_vocab(vocab_path)
if hotword:
print(f"已加载专业词汇：{hotword}")
else:
print("未找到词汇表文件，使用默认识别模式")
print("\n正在加载模型...")
model = AutoModel(
model="FunAudioLLM/Fun-ASR-Nano-2512",
hub="hf",
trust_remote_code=True,
vad_model="funasr/fsmn-vad",
vad_kwargs={"max_single_segment_time": 30000},
spk_model="funasr/campplus",
device="cuda:0",
)
print("模型加载完成")
audio_extensions = ('.mp3', '.wav', '.flac', '.ogg', '.m4a', '.MP3', '.WAV', '.FLAC')
audio_files = [f for f in os.listdir(audio_dir) if f.endswith(audio_extensions)]
if not audio_files:
print("未找到音频文件")
return
print(f"\n找到 {len(audio_files)} 个音频文件")
success_count = 0
fail_count = 0
for audio_file in audio_files:
audio_path = os.path.join(audio_dir, audio_file)
if transcribe_audio(audio_path, output_dir, hotword, model):
success_count += 1
else:
fail_count += 1Expected behavior
no error
Error logs
<html>
<body>
<!--StartFragment-->
2441.9s | 18781 |
-- | -- | --
2441.9s | 18782 |
2441.9s | 18783 | {'load_data': '0.000', 'extract_feat': '0.007', 'forward': '0.652', 'batch_size': '1', 'rtf': '0.247'}, : 55%\|�[34m█████▍ �[0m\| 56/102 [00:33<00:28, 1.64it/s]
2441.9s | 18784 | {'load_data': '0.000', 'extract_feat': '0.007', 'forward': '0.652', 'batch_size': '1', 'rtf': '0.247'}, : 55%\|�[34m█████▍ �[0m\| 56/102 [00:33<00:28, 1.64it/s]�[A
2442.6s | 18785 | �[A {'load_data': '0.000', 'extract_feat': '0.007', 'forward': '0.652', 'batch_size': '1', 'rtf': '0.247'}, : 55%\|�[34m█████▍ �[0m\| 56/102 [00:34<00:27, 1.65it/s]
2442.6s | 18786 | {'load_data': '0.000', 'extract_feat': '0.007', 'forward': '0.652', 'batch_size': '1', 'rtf': '0.247'}, : 55%\|�[34m█████▍ �[0m\| 56/102 [00:34<00:27, 1.65it/s]
2442.6s | 18787 | 0%\|�[31m �[0m\| 0/1 [06:48<?, ?it/s]
2442.6s | 18788 |
2442.7s | 18789 | 转录失败: Encountered text corresponding to disallowed special token '<\|no\|>'.
2442.7s | 18790 | If you want this text to be encoded as a special token, pass it to `allowed_special`, e.g. `allowed_special={'<\|no\|>', ...}`.
2442.7s | 18791 | If you want this text to be encoded as normal text, disable the check for this token by passing `disallowed_special=(enc.special_tokens_set - {'<\|no\|>'})`.
2442.7s | 18792 | To disable this check for all special tokens, pass `disallowed_special=()`.
2442.7s | 18793 |
<br class="Apple-interchange-newline"><!--EndFragment-->
</body>
</html>
Environment
- OS: Linux 6.12.90+
- Python version: 3.12.13
- FunASR version: 1.3.14
- ModelScope version: 1.38.0
- PyTorch / torchaudio version: 2.10.0+cu128 / 2.10.0+cu128
- Install method (pip, source, Docker): pip
- Device (cuda, cpu, mps): cuda
- GPU model: Tesla T4
- CUDA/cuDNN version: 12.8 / 91002
- Docker image tag, if used: gcr.io/kaggle-gpu-images/python@sha256:37c64f7dd9c54116ecd1bcc88817c5469b88387388fade02bfa8bf3fc647d461
Audio details
If the audio cannot be shared, please describe:
Duration 3h 13m 38s（11,617s）
Sample rate 16,000 Hz
Format MP3 (MPEG audio layer 3)
Bit rate 64 kbps
Channels 2 (stereo)
Sample format 32-bit float planar
File size 88.6 MB
🐛 Bug
ASR transcription failed with special token validation error during batch dataset inference, while most samples run normally with stable RTF performance.
To Reproduce
code: https://www.kaggle.com/code/liuweiq/funasr-recog-script?scriptVersionId=333049663
Code sample
Expected behavior
no error
Error logs
Environment
Audio details
If the audio cannot be shared, please describe:
Duration 3h 13m 38s（11,617s）
Sample rate 16,000 Hz
Format MP3 (MPEG audio layer 3)
Bit rate 64 kbps
Channels 2 (stereo)
Sample format 32-bit float planar
File size 88.6 MB