fishaudio/fish-speech · error · ValueError

Number of prompt text ({len(prompt_text)}) and prompt audio

Error message

Number of prompt text ({len(prompt_text)}) and prompt audio ({len(prompt_audio)}) should be the same

What it means

Raised by the text2semantic inference CLI when the number of prompt texts and prompt audio files passed for voice cloning differ. The model pairs each reference text with its corresponding reference audio, so the lists must have equal length.

Source

Thrown at fish_speech/models/text2semantic/inference.py:871

    seed: int,
    half: bool,
    iterative_prompt: bool,
    chunk_length: int,
    output_dir: Path,
) -> None:
    os.makedirs(output_dir, exist_ok=True)
    precision = torch.half if half else torch.bfloat16

    if prompt_text and not prompt_audio and not prompt_tokens:
        raise ValueError(
            "--prompt-text requires either --prompt-audio or --prompt-tokens"
        )
    if prompt_text and prompt_tokens and len(prompt_text) != len(prompt_tokens):
        raise ValueError(
            f"Number of prompt text ({len(prompt_text)}) and prompt tokens ({len(prompt_tokens)}) should be the same"
        )
    if prompt_text and prompt_audio and len(prompt_text) != len(prompt_audio):
        raise ValueError(
            f"Number of prompt text ({len(prompt_text)}) and prompt audio ({len(prompt_audio)}) should be the same"
        )

    logger.info("Loading model ...")
    t0 = time.time()
    model, decode_one_token = init_model(
        checkpoint_path, device, precision, compile=compile
    )
    with torch.device(device):
        model.setup_caches(
            max_batch_size=1,
            max_seq_len=model.config.max_seq_len,
            dtype=next(model.parameters()).dtype,
        )
    if torch.cuda.is_available():
        torch.cuda.synchronize()

    logger.info(f"Time to load model: {time.time() - t0:.02f} seconds")

View on GitHub (pinned to befe400174)

Solutions

1. Count --prompt-text and --prompt-audio arguments and make them equal (1:1 pairing, in the same order)
2. Omit both to synthesize without voice cloning
3. Ensure empty strings are not accidentally passed as extra entries (e.g. trailing commas in shell vars)

Example fix

# before
--prompt-audio a.wav b.wav --prompt-text "one"
# after
--prompt-audio a.wav b.wav --prompt-text "one" "two"

assert len(prompt_text or []) == len(prompt_audio or []), f"{len(prompt_text)} texts vs {len(prompt_audio)} audios"

When it happens

Trigger: Calling inference.py with --prompt-text values whose count differs from the number of --prompt-audio files (e.g. 2 texts and 1 audio).

Common situations: Constructing a voice-cloning command by hand, or a wrapper script that appends an extra reference audio without a matching transcript.

Related errors

• Directory {path} does not exist.
• --reference_audio and --reference_text must be given the sam
• Reference audio file not found: {ref_audio}
• Invalid quantization mode {mode} needs to be one of [int8, i
• Text is too long, max length is {app_state.max_text_length}

AI-assisted analysis of fishaudio/fish-speech@befe400174 (2026-08-27). Data as JSON: /api/errors/61d31119c5bbae05. Report an issue: GitHub.