fishaudio/fish-speech · error · HTTPException

{str(result.error)}

Error message

{str(result.error)}

What it means

The streaming TTS wrapper maps internal 'error' results from the inference pipeline to HTTP 500 with the underlying error string as content. It is a pass-through of failures inside model inference (loading, encoding, generation).

Source

Thrown at tools/server/inference.py:25

from fish_speech.utils.schema import ServeTTSRequest

AMPLITUDE = 32768  # Needs an explaination


def inference_wrapper(req: ServeTTSRequest, engine: TTSInferenceEngine):
    """
    Wrapper for the inference function.
    Used in the API server.
    """
    count = 0
    for result in engine.inference(req):
        match result.code:
            case "header":
                if isinstance(result.audio, tuple):
                    yield result.audio[1]

            case "error":
                raise HTTPException(
                    HTTPStatus.INTERNAL_SERVER_ERROR,
                    content=str(result.error),
                )

            case "segment":
                count += 1
                if isinstance(result.audio, tuple):
                    yield (result.audio[1] * AMPLITUDE).astype(np.int16).tobytes()

            case "final":
                count += 1
                if isinstance(result.audio, tuple):
                    yield result.audio[1]
                return None  # Stop the generator

    if count == 0:
        raise HTTPException(
            HTTPStatus.INTERNAL_SERVER_ERROR,

View on GitHub (pinned to befe400174)

Solutions

1. Read the returned content string and the server logs — it names the real underlying error
2. Fix the root cause (model paths, memory, reference audio format)
3. Reduce text length / batch to rule out OOM

try:
    for chunk in stream_tts(client, req):
        ...
except HTTPError as e:
    if e.response.status_code == 500:
        log.error(e.response.text)  # underlying error string

When it happens

Trigger: Any exception raised during the serve pipeline — bad model path, unsupported request combination, OOM during generation — surfaces here as a 500 with the original message.

Common situations: Server started with mismatched model configs, corrupted audio references, CUDA OOM on long texts, or invalid parameters that pass request validation but break inference.

Related errors

• No audio generated, please check the input text.
• Failed to encode audio
• Failed to decode tokens to audio
• Expected GenerateResponse, got {type(wrapped_result.response
• Invalid token

AI-assisted analysis of fishaudio/fish-speech@befe400174 (2026-08-27). Data as JSON: /api/errors/2758f8e3adc044cf. Report an issue: GitHub.