microsoft/VibeVoice · error · ValueError

Unsupported audio type: {type(audio)}

Error message

Unsupported audio type: {type(audio)}

What it means

`save_audio()` only accepts `torch.Tensor`, `np.ndarray`, or a `list` of tensors/arrays (see the Union in its signature at vibevoice/processor/vibevoice_tokenizer_processor.py:284). Any other type falls through to `raise ValueError(f"Unsupported audio type: {type(audio)}")` before any file I/O happens. It is an input-contract error raised by explicit type dispatch, not by serialization.

Source

Thrown at vibevoice/processor/vibevoice_tokenizer_processor.py:331

            raise ImportError(
                "soundfile is required to save audio files. "
                "Install it with: pip install soundfile"
            )
        
        # Ensure audio is in the right format
        if isinstance(audio, torch.Tensor):
            # Convert PyTorch tensor to numpy
            audio_np = audio.float().detach().cpu().numpy()
        elif isinstance(audio, np.ndarray):
            audio_np = audio
        elif isinstance(audio, list):
            # Handle list of tensors or arrays
            if all(isinstance(a, torch.Tensor) for a in audio):
                audio_np = [a.float().detach().cpu().numpy() for a in audio]
            else:
                audio_np = audio
        else:
            raise ValueError(f"Unsupported audio type: {type(audio)}")
        
        saved_paths = []
        
        # Handle based on shape or type
        if isinstance(audio_np, list):
            # Multiple separate audios to save
            output_dir = output_path
            
            # Ensure output directory exists
            os.makedirs(output_dir, exist_ok=True)
            
            # Save each audio
            for i, audio_item in enumerate(audio_np):
                audio_item = self._prepare_audio_for_save(audio_item, normalize)
                file_path = os.path.join(output_dir, f"{batch_prefix}{i}.wav")
                sf.write(file_path, audio_item, sampling_rate)
                saved_paths.append(file_path)
                

View on GitHub (pinned to 94da20d98b)

Solutions

1. Convert the value to a supported type before saving: wrap tuples with `list(...)`, convert non-torch arrays with `torch.as_tensor(...)` or `np.asarray(...)`.
2. If you meant to load audio, use the processor's audio-loading API (or soundfile/librosa) — `save_audio` only writes.
3. For mixed lists (tensors + numpy), normalize first: `[a.numpy() if isinstance(a, torch.Tensor) else a for a in audio]`.

Example fix

# before
tracks = tuple(w.cpu() for w in wavs)  # tuple -> ValueError
processor.save_audio(tracks, "out_dir")

# after
tracks = [w.detach().cpu().numpy() for w in wavs]  # list of arrays
processor.save_audio(tracks, "out_dir")

import torch, numpy as np

def as_saveable(audio):
    if isinstance(audio, (torch.Tensor, np.ndarray)):
        return audio
    if isinstance(audio, (list, tuple)):
        return [a.detach().cpu().numpy() if isinstance(a, torch.Tensor) else a for a in audio]
    raise TypeError(f"Cannot save audio of type {type(audio)!r}")

processor.save_audio(as_saveable(wav), "out.wav")

def is_saveable_audio(audio) -> bool:
    import torch
    if isinstance(audio, (torch.Tensor, np.ndarray)):
        return True
    return isinstance(audio, (list, tuple)) and all(
        isinstance(a, (torch.Tensor, np.ndarray)) for a in audio
    )

try:
    processor.save_audio(wav, out)
except ValueError as e:
    if "Unsupported audio type" in str(e):
        wav = list(wav) if isinstance(wav, tuple) else wav
        processor.save_audio(wav, out)
    else:
        raise

When it happens

Trigger: Passing a Python tuple of tensors, a generator, a string path to a wav file, a JAX/TF tensor, or a 0-d/odd object to `save_audio()`. Also passing a nested list of lists, since only a flat list of torch.Tensors is converted element-wise.

Common situations: Pipeline code that collects outputs into tuples instead of lists, feeding a model output after `.cpu()` chained onto a tuple, or handing `save_audio` a file path expecting it to read audio (it only writes).

Related errors

• soundfile is required to save audio files. Install it with:
• Sample index {sample_idx} exceeds batch size {self.batch_siz
• load_audio_bytes_use_ffmpeg requires resample=True
• Audio input is required for ASR processing
• Unexpected audio shape: {audio.shape}

AI-assisted analysis of microsoft/VibeVoice@94da20d98b (2026-08-15). Data as JSON: /api/errors/66473767d15847d7. Report an issue: GitHub.