DualCodec — Hindi, 25 Hz
DualCodec neural speech codec fine-tuned on Hindi, operating at 25 Hz.
Part of the TinyAya codec study, which asked whether fine-tuning a neural audio codec on a low-resource language improves reconstruction over the stock multilingual checkpoint — the same question that bounds the S2ST model, whose audio quality is capped by its frozen decoder.
Weights: model.safetensors, model_1.safetensors. Produced by
codec-finetuning, which fine-tunes Mimi, DualCodec and
Kanade on Turkish and Hindi across 8 optimizers with W&B Bayesian sweeps and
bootstrap error bars.
Upstream DualCodec licence terms apply.
Code
| repo | what it does |
|---|---|
codec-finetuning |
fine-tunes Mimi / DualCodec / Kanade on Turkish + Hindi |
Project
TinyAya Stage 2 — Turkish⇄Hindi speech-to-speech translation with a text inner-monologue: a LoRA-adapted Cohere2 backbone driving a frozen Moshi depth decoder over Mimi codes.
The v0.3 run covered 76,250 steps / 2.07 epochs on a Cloud TPU v6e-16 (best val composite 2.8199 @ step 76,000). Read honestly: the text inner-monologue learns to translate (free-run chrF++ ~25.7 / 25.1), while intelligible audio synthesis remains the frontier (ASR-chrF++ 3.7 / 9.6 against a 92.1 / 86.6 ground-truth-audio ceiling) — bounded by the frozen depth decoder, not by translation understanding.
- Results: v0.3 evaluation report
- Training run: W&B
xzcb60bl· emergence report - Blog: Adapting Moshi for Low-Resource Speech Translation
Compute for the v0.3 run was provided by Google's TPU Research Cloud (TRC).