unslothai/unsloth · error · ValueError

No valid audio codes found after START_OF_SPEECH token

Error message

No valid audio codes found after START_OF_SPEECH token

What it means

Raised while decoding Orpheus-style SNAC audio codes: after cropping generated token ids at the START_OF_SPEECH token (128257), stripping EOS (128258) and truncating to a multiple of 7, the remaining row is empty. The model produced no audio frames — only markers/EOS or an immediate end after speech start.

Source

Thrown at studio/backend/core/inference/audio_codecs.py:163

        Returns (wav_bytes, 24000).
        """
        # Find START_OF_SPEECH token (128257)
        token_indices = (generated_ids == 128257).nonzero(as_tuple = True)
        if len(token_indices[1]) > 0:
            cropped = generated_ids[:, token_indices[1][-1] + 1 :]
        else:
            # Fall back to the entire output if the marker is missing
            logger.warning("No START_OF_SPEECH token (128257) found — using full generated output")
            cropped = generated_ids
        row = cropped[0]

        # Remove EOS tokens (128258)
        row = row[row != 128258]

        # Trim to multiple of 7
        row = row[: (len(row) // 7) * 7]
        if len(row) == 0:
            raise ValueError("No valid audio codes found after START_OF_SPEECH token")

        codes = [t.item() - 128266 for t in row]

        # Redistribute into 3 SNAC layers (7 codes per frame → 1+2+4)
        layer_1, layer_2, layer_3 = [], [], []
        for i in range(len(codes) // 7):
            layer_1.append(codes[7 * i])
            layer_2.append(codes[7 * i + 1] - 4096)
            layer_3.append(codes[7 * i + 2] - 8192)
            layer_3.append(codes[7 * i + 3] - 12288)
            layer_2.append(codes[7 * i + 4] - 16384)
            layer_3.append(codes[7 * i + 5] - 20480)
            layer_3.append(codes[7 * i + 6] - 24576)

        snac_codes = [
            torch.tensor(layer).unsqueeze(0).to(device) for layer in [layer_1, layer_2, layer_3]
        ]

View on GitHub (pinned to 203007d190)

Solutions

1. Retry generation with different sampling (lower temperature, adjust repetition penalty) — degenerate continuations are often stochastic
2. Increase max_new_tokens so at least one 7-code frame can be produced
3. Verify non-empty input text and a valid speaker/voice reference
4. If it consistently fails, the checkpoint or tokenizer mapping is wrong for the audio head — reload the correct TTS model files

def has_audio_frames(generated_ids, sos=128257, eos=128258) -> bool:
    ids = generated_ids[0]
    idx = (ids == sos).nonzero()
    row = ids[idx[-1].item()+1:] if len(idx) else ids
    row = row[row != eos]
    return len(row) >= 7

try:
    wav, sr = decode_snac(generated_ids)
except ValueError as e:
    if "No valid audio codes" in str(e):
        regenerate_with(seed=None, temperature=lower_temp)

When it happens

Trigger: The TTS model generates EOS immediately after START_OF_SPEECH (silent/degenerate continuation); audio-token head misconfigured so no valid code ids in range appear; generation parameters (temperature, repetition penalty, max tokens) truncate output before any 7-code frame; a fallback path used the whole output when the SOS marker was absent and it contained no usable codes.

Common situations: Wrong or mismatched voice/prompt tokens for the Orpheus model; max_new_tokens set too low; quantized/GGUF TTS model dropping audio tokens; empty or whitespace input text.

Related errors

• No bicodec_semantic tokens found in generated output
• No DAC code tokens (c1/c2) found in generated output
• Model {self.active_model_name} is not an audio model
• Audio generation cancelled
• Unknown audio_type: {audio_type}

AI-assisted analysis of unslothai/unsloth@203007d190 (2026-08-15). Data as JSON: /api/errors/0a32453a23517156. Report an issue: GitHub.