microsoft/VibeVoice · error · ValueError

Unsupported tokenizer type for {language_model_pretrained_na

Error message

Unsupported tokenizer type for {language_model_pretrained_name}

What it means

VibeVoiceASRProcessor.from_pretrained builds its text tokenizer from `language_model_pretrained_name`. The only supported family is Qwen (detected by a case-insensitive 'qwen' substring in the repo name); any other name — Llama, Gemma, a local fine-tune directory without 'qwen' in the path — raises this ValueError before the audio processor is built.

Source

Thrown at vibevoice/processor/vibevoice_asr_processor.py:147

                logger.warning(f"Could not load preprocessor_config.json: {e}")
                logger.warning("Using default configuration")
        
        # Extract parameters
        speech_tok_compress_ratio = config.get("speech_tok_compress_ratio", 3200)
        target_sample_rate = config.get("target_sample_rate", 24000)
        normalize_audio = config.get("normalize_audio", True)
        
        # Load tokenizer
        language_model_pretrained_name = config.get("language_model_pretrained_name", None) or kwargs.pop("language_model_pretrained_name", "Qwen/Qwen2.5-1.5B")
        logger.info(f"Loading tokenizer from {language_model_pretrained_name}")
        
        if 'qwen' in language_model_pretrained_name.lower():
            tokenizer = VibeVoiceASRTextTokenizerFast.from_pretrained(
                language_model_pretrained_name,
                **kwargs
            )
        else:
            raise ValueError(f"Unsupported tokenizer type for {language_model_pretrained_name}")
        
        # Load audio processor
        audio_processor = VibeVoiceTokenizerProcessor(
            sampling_rate=target_sample_rate,
            normalize_audio=normalize_audio,
            target_dB_FS=config.get("target_dB_FS", -25),
            eps=config.get("eps", 1e-6),
        )
        
        return cls(
            tokenizer=tokenizer,
            audio_processor=audio_processor,
            speech_tok_compress_ratio=speech_tok_compress_ratio,
            target_sample_rate=target_sample_rate,
            normalize_audio=normalize_audio,
        )
    
    def save_pretrained(self, save_directory: Union[str, os.PathLike], **kwargs):

View on GitHub (pinned to 94da20d98b)

Solutions

Use a Qwen-family checkpoint id such as 'Qwen/Qwen2.5-1.5B' (the default) for language_model_pretrained_name.
If using a local fine-tune, ensure the directory/path string still contains 'qwen' (case-insensitive), or rename it accordingly.
For other model families, subclass the processor and add a tokenizer branch instead of relying on this loader.

Example fix

# before
processor = VibeVoiceASRProcessor.from_pretrained('/checkpoints/my-asr-lm')  # no 'qwen' in path

# after
processor = VibeVoiceASRProcessor.from_pretrained(
    ..., language_model_pretrained_name='Qwen/Qwen2.5-1.5B'
)

Defensive patterns

Strategy: validation

Validate before calling

name = config.get('language_model_pretrained_name') or 'Qwen/Qwen2.5-1.5B'
if 'qwen' not in name.lower():
    raise ValueError(f'ASR processor requires a Qwen tokenizer, got {name!r}')
processor = VibeVoiceASRProcessor.from_pretrained(..., language_model_pretrained_name=name)

Type guard

def is_qwen_tokenizer_name(name: str) -> bool:
    return isinstance(name, str) and 'qwen' in name.lower()

Try / catch

try:
    processor = VibeVoiceASRProcessor.from_pretrained(model_path)
except ValueError as e:
    if 'Unsupported tokenizer type' in str(e):
        processor = VibeVoiceASRProcessor.from_pretrained(
            model_path, language_model_pretrained_name='Qwen/Qwen2.5-1.5B')
    else:
        raise

Prevention

Pin language_model_pretrained_name to a known Qwen repo id in your config.
Keep 'qwen' in local checkpoint directory names so the substring check matches.
Fail early with your own clearer message when the name is non-Qwen.

When it happens

Trigger: Passing a config or kwarg language_model_pretrained_name that does not contain 'qwen', e.g. 'meta-llama/Llama-3.1-8B', 'google/gemma-2-2b', or a local path like '/models/my-finetune' even when the underlying model is Qwen-based.

Common situations: Pointing at a locally saved copy of Qwen whose directory name dropped the 'qwen' token; attempting to swap the LM backbone to Llama/Gemma (the error message's sibling in vibevoice_processor.py even claims Llama/Gemma are supported — they are not wired up here); typos in the repo id.

Related errors

AI-assisted analysis of microsoft/VibeVoice@94da20d98b (2026-08-15). Data as JSON: /api/errors/104059ac09b94ced. Report an issue: GitHub.