MOSS-12to32-Enhancer

A small causal Qwen3 model that predicts 32 MOSS v2 audio-codebook indices from the 12 indices of the current frame and up to ten previously generated HQ frames. It is trained from scratch on German, English, Spanish and French speech.

This is an experimental enhancement model. Validation cross entropy measures token prediction under teacher forcing. It does not establish an improvement in perceived audio quality. Use the generated-history listening evaluation to judge speech content, speaker identity, artifacts and long-utterance stability.

Current release status

Training state: complete. Progress snapshot: 100.00% of one epoch. Root inference weights: step 64,699, selected by held-out extra-20-codebook cross entropy 5.818123.

Listening reports:

• complete · Step 12,940: 100/100 snippets / 200 input cases (download the HTML and open locally).
• partial · Step 12,940: 1/100 snippets / 2 input cases (download the HTML and open locally).
• complete · Step 12,940: 100/100 snippets / 200 input cases (download the HTML and open locally).
• complete · Step 22,645: 4/4 snippets / 8 input cases (download the HTML and open locally).
  • Same report with DNSMOS and reconstruction metrics
• complete · Step 48,525: 4/4 snippets / 8 input cases (download the HTML and open locally).
  • Same report with DNSMOS and reconstruction metrics
• complete · Step 64,699: 100/100 snippets / 200 input cases (download the HTML and open locally).
• complete · Step 64,699: 50/50 snippets / 100 input cases (download the HTML and open locally).
  • Same report with DNSMOS and reconstruction metrics

This status is refreshed on checkpoint/evaluation commits. Training progress is a snapshot, not a live counter.

Completed SAPO / GRPO-family release

The completed group-relative RL pilot is available on the rl-sapo-v1 branch. Its English README provides executable token inference, exact adapter/base loading, merged checkpoint inference, recorded-rollout optimizer replay and fresh online training instructions. The selected best checkpoint is group 52; the final group 104 and every intermediate checkpoint are retained separately. The objective is SAPO with a composite 55% audio-quality / 45% token-fidelity reward, rank-64 adapters and frozen-base KL plus supervised anchors. The pilot completed 104 groups, eight candidates each and 208 optimizer updates. Its small fixed evaluation does not establish broad quality improvements.

All ranked candidates and targets are stored in the linked token dataset. New supervised wideband tokens are in expansion_v2/1000h; the 1,000 h directory name is a target, and its manifest states the actual published quantity. Those expansion rows were not used in the completed pilot.

What goes in and out

This increases the number of residual codebooks at the same frame rate. There is no temporal interpolation. Ten history frames are 0.8 seconds. Humaneness-Voice-Small and any text language model weights are not needed. Talker-generated tokens can be passed directly when they use this same codec revision and codebook ordering; that input distribution has not yet been validated by the codec-encoded speech evaluation.

The dataset stores arrays as [frames, codebooks]. The codec's decode interface uses [codebooks, batch, frames]. An integer value has a different meaning in each codebook; do not flatten the inputs into a single shared 1024-way vocabulary.

Architecture