fishaudio/fish-speech · error · HTTPException

Failed to decode tokens to audio

Error message

Failed to decode tokens to audio

What it means

The /v1/vqgan/decode endpoint converts token tensors back to audio; any exception during decoding is logged and converted to HTTP 500 'Failed to decode tokens to audio'. The token payload or model state is at fault.

Solutions

1. Check server logs for the underlying exception details
2. Ensure tokens match the shape produced by /v1/vqgan/encode (list of codebook arrays)
3. Regenerate tokens with the same model version the server loaded

import numpy as np
tokens = [np.asarray(t, dtype=np.int64) for t in tokens]
assert all(t.ndim == 1 for t in tokens), "each token array must be 1-D"
assert all(t.min() >= 0 and t.max() < num_codes for t in tokens)

if resp.status_code == 500:
    raise RuntimeError("token payload malformed or from a different VQ model version")

When it happens

Trigger: Sending tokens with wrong shape/dtype/values (e.g. flat lists instead of [codes, t] shaped structures, out-of-range codebook indices) to /v1/vqgan/decode.

Common situations: Clients serializing tokens incorrectly (tolist on the wrong axis), using tokens from a different VQ model version, or truncated msgpack payloads.

Understand the failure class

• Parsing and encoding errors: unexpected token, malformed input — why parsers reject input and how to find the real culprit.

Related errors

• Failed to encode audio
• No audio generated, please check the input text.
• {str(result.error)}
• Invalid token
• Unsupported Media Type

AI-assisted analysis of fishaudio/fish-speech@befe400174 (2026-08-27). Data as JSON: /api/errors/ea28175629a91f5f. Report an issue: GitHub.

Appendix: source

Thrown at tools/server/views.py:141

        decoder_model = model_manager.decoder_model

        # Decode the audio
        tokens = [torch.tensor(token, dtype=torch.int) for token in req.tokens]
        start_time = time.time()
        audios = batch_vqgan_decode(decoder_model, tokens)
        logger.info(
            f"[EXEC] VQGAN decode time: {(time.time() - start_time) * 1000:.2f}ms"
        )
        audios = [audio.astype(np.float16).tobytes() for audio in audios]

        # Return the response
        return ormsgpack.packb(
            ServeVQGANDecodeResponse(audios=audios),
            option=ormsgpack.OPT_SERIALIZE_PYDANTIC,
        )
    except Exception as e:
        logger.error(f"Error in VQGAN decode: {e}", exc_info=True)
        raise HTTPException(
            HTTPStatus.INTERNAL_SERVER_ERROR, content="Failed to decode tokens to audio"
        )


@routes.http.post("/v1/tts")
async def tts(req: Annotated[ServeTTSRequest, Body(exclusive=True)]):
    """
    Generate speech from text using TTS model.
    """
    try:
        # Get the model from the app
        app_state = request.app.state
        model_manager: ModelManager = app_state.model_manager
        engine = model_manager.tts_inference_engine
        sample_rate = engine.decoder_model.sample_rate

        # Check if the text is too long
        if app_state.max_text_length > 0 and len(req.text) > app_state.max_text_length:

View on GitHub (pinned to befe400174)