WER of token2text

[anshkumar]

I'm trying to implement token2text for my TTS. I wanted to know how much WER were you getting with your token2text. Can I see some training loss graph and some results of token2text ?

[aluminumbox]

我是CosyVoice自动答疑助手（automated bot reply — 我是机器人，非人工回复）。

A few pointers on your question:

• CosyVoice's speech tokenizer (S3Tokenizer, derived from the SenseVoice ASR encoder) produces supervised semantic speech tokens, not text tokens. The released stack in this repo is: S3Tokenizer (audio → speech tokens) → LLM (text → speech tokens) → flow-matching + HiFT decoder (tokens → waveform). A standalone "token → text" head is not part of the released pipeline, so the repo does not publish WER numbers, training loss curves, or checkpoints for it.
• The CosyVoice / CosyVoice2 / CosyVoice3 papers discuss the ASR-supervised nature of the semantic tokens (i.e., they are trained so that a small ASR decoder can recover text with low WER on the training languages) but the exact WER of any downstream token2text head will depend on your decoder architecture, training data, and test set.
• If you want to build one yourself, the usual recipe is:
  1. Run S3Tokenizer over a labelled set (e.g. LibriSpeech / AISHELL / your own domain data) to get (tokens, text) pairs.
  2. Train a lightweight decoder (Transformer / CTC / RNN-T) on those pairs.
  3. Evaluate WER on the standard test splits.
• For authoritative and up-to-date details please refer to the latest README, examples/, and the CosyVoice papers in this repo — those are the source of truth if anything above drifts from the current code.

This question has been logged for the maintainers in case they'd like to share any internal ablation on token2text WER. Hope this helps you get started.