Qwen3-TTS multi-codebook tokenizer (Transformers conversion)
Conversion of the Qwen3-TTS multi-codebook (12Hz) speech tokenizer to the Transformers implementation added in huggingface/transformers#44517. Stored in float32, matching the original.
Converted from Qwen/Qwen3-TTS-Tokenizer-12Hz with the conversion script that
ships with the model. The tensors are unchanged: conversion only renames state dict keys to match the
Transformers module layout, fuses the code predictor's per-group lm_head layers into one projection,
and rewrites the configuration.
This checkpoint exists so the integration tests have something to fetch while the model is under review. Once merged, a Transformers-compatible checkpoint is expected to be hosted by the Qwen team.
Usage
from transformers import Qwen3TTSTokenizerMultiCodebookModel
model = Qwen3TTSTokenizerMultiCodebookModel.from_pretrained("shahvandit/qwen3-tts-tokenizer-multi-codebook-hf")
License
Apache 2.0, inherited from the original model.