MOSS-TTS Decode Tokens

Most of a MOSS-TTS workflow runs in a space you can't hear. The model thinks in audio codes at 12.5 frames per second, and normally they stay hidden inside the generate nodes. MOSS-TTS Decode Tokens drags them back out into a normal ComfyUI AUDIO clip.

Which sounds like a novelty until you've spent an hour fighting a clone that comes out as gibberish. This is the tool that tells you where the mistake is. It's also how you check a saved base voice, or the exact reference window a Concat Tokens node just built, before you spend a batch on it. Encode/decode/concat/save/load is a complete token API, and decode is the ear on it.

How it works

It's the inverse of MOSS-TTS Encode Tokens: codes go back through the model's own vocoder and come out as PCM. There is no sampling involved - no temperature, no top-k, no seed. The same codes always produce the same audio, every time. That's what makes it a diagnostic and not a generation: if two decodes of the same file differ, something upstream changed the file.

The vocoder lives in the model bundle, so the model is a required input, and the sample rate follows it - 48 kHz for the default Local-Transformer, 24 kHz for the 8B. n_vq (the number of codebooks per frame) is validated, so codes from the wrong variant stop with a clear error instead of quietly making static. Verify that error if you see it: it almost always means you encoded on one model and switched the loader to the other.

Inputs and outputs

• moss_model - from MOSS-TTS Load Model. Same variant that encoded or generated the codes.
• tokens - anything MOSS_TOKENS: the output of Encode Tokens, Concat Tokens, Load Tokens, or the tokens output of Speak / Voice Clone / Voice Continue.
• return_stereo (optional, default true) - keeps the codec's native stereo, identical to what every generate node in this pack outputs. Set it false and the channels are averaged into mono: smaller, but no longer bit-identical to the generated audio. Default-model only - the 8B's codec is mono and ignores the switch.

Outputs are audio (AUDIO at the model's native rate, ready for PreviewAudio / SaveAudio) and frames (INT - divide by 12.5 for seconds).

That return_stereo detail is worth caring about if you're feeding the result into anything that compares or mixes audio: a decoded tokens output with the default left on lines up sample-for-sample with the audio the same node emitted.

What to actually use it for

Three things, in order of how often you'll want them:

Audition the reference before a long run. Hang a decode off the token stream your chain actually assembled - the saved base voice, the concatenated sliding window - and preview it. A reference that sounds wrong here will clone wrong. Cheaper to find out now.

Find out what's in a .moss_tokens.pt file. Tokens are opaque; MOSS-TTS Save Tokens gives you a path and no guarantee the file holds the voice you think it does. Load Tokens → Decode Tokens → PreviewAudio answers it in one codec pass.

Explain the 80 ms. Voice Continue trims its first frame (default head_trim_frames = 1) off the audio output, but its tokens output is untrimmed - that's deliberate, MOSS needs the codes whole as the next prefix. So decoding those tokens returns audio that is exactly head_trim_frames longer than the audio beside it. If you were wondering why a decode doesn't match the file frame for frame, that's why. Set head_trim_frames = 0 if you need a 1:1 match.

Installing it

ComfyUI-Manager: search MOSS-TTS 1.5 by enrico. Not the other packs with "MOSS-TTS" in the title. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/eehrich/ComfyUI-MOSS-TTS-1.5.git

Restart ComfyUI. No pip dependencies - the pack's pyproject.toml intentionally declares none, leaning on the torch / torchaudio / transformers your ComfyUI already has. First load of MOSS-TTS Load Model still fetches ~9 GB of model plus ~8.5 GB of audio tokenizer unless you already have them in ComfyUI/models/moss_tts.

Troubleshooting

• "codes were made by a different model" style error - the n_vq validation doing its job. Re-encode with the model you're generating on.
• Mono output on the 8B despite return_stereo = true - expected. The 8B is mono end to end; the switch is documented for the default model only.
• Nothing plays - check the sample rate the preview is being asked for. A 48 kHz decode and a 24 kHz decode are different files, and some downstream audio tools in ComfyUI care.
• Loading the model fails with non-default argument 'sampling_rate' follows default argument - you're on transformers 5.4.0 or 5.5.0. python -m pip install -U "transformers>=5.5.1" inside ComfyUI's Python environment, or hand-patch the tokenizer config as the pack README describes.

Worth saying plainly: this is a debugging and QC node. If your pipeline runs clean, you may never wire it. The moment a voice goes wrong, it's the first thing I'd reach for - it costs one codec pass and it removes all the guessing from a chain that is otherwise invisible.