quantize HuBERT vectors to bark semantic tokens

The last piece of the voice-cloning puzzle. bark_hubert_quantize takes the continuous HuBERT vectors from bark_hubert_encode and runs them through the quantizer loaded by load_bark_hubert_quantizer, producing Bark's semantic tokens. That's the moment a reference voice officially becomes something you can feed into Bark's generation chain - or save into a speaker .npz.

How it works

Inside the quantizer, each 768-dimensional vector passes through a two-layer LSTM followed by a linear layer that scores all 10,000 semantic tokens, and the node takes the argmax - the single most likely token for each vector. It's a hard quantization: continuous "how does this sound" becomes a discrete index Bark's semantic model understands. That's also why it's fast and why it's one-way - you can't recover the vectors from the tokens, and you don't need to.

Inputs & outputs

• quantizer_model - from load_bark_hubert_quantizer.
• vectors - from bark_hubert_encode.

The output, semantic_tokens (BarkSemanticTokens), is the semantic prompt for a speaker file: wire it into save_speaker_npz's semantic_prompt input alongside the coarse and fine prompts from the encodec chain, and you've got a complete .npz that load_speaker_npz can bring back later.

Install & the fine print

Pack install, standard:

cd ComfyUI/custom_nodes
git clone https://github.com/gitmylo/ComfyUI-audio-nodes

(or ComfyUI Manager → "Audio nodes"), restart.

Two practical notes. First, the layer/quantizer pairing is everything: bark_hubert_encode defaults to layer 9, and you want the quantizer that was trained on layer 9 - if you grabbed quantifier_hubert_base_ls960_14.pth from the loader dropdown thinking "newer default = better," the mismatch will quietly degrade your cloning. Second, because quantization is argmax (greedy, no temperature), it's deterministic - the same clip always yields the same tokens, which is what you want when building a speaker file you'll reuse. That determinism is also the honest limit of the whole approach: the speaker is only as good as the reference clip, and HuBERT-to-token is a lossy translation. Clean 5–10 second clip, one speaker, no music, and you'll get a usable voice anchor; anything messier and you'll be fighting the source material the whole way down.