Overview
Universal Audio Tokenizer is a WeChat AI audio tokenizer designed as a compact, single-codebook interface for audio language models. It combines linguistic alignment with representations of general audio, including speech and other sound events. (github.com)
Capabilities and uses
- Encodes audio clips into discrete tokens and can reconstruct waveforms from those tokens using the supplied inference example.
- Uses Semantic-Acoustic Primitives supervision and a Semantic-Acoustic Equilibrium mechanism to combine linguistic and acoustic information.
- Intended for integration into audio-LLM systems for audio understanding, speech reconstruction and text-to-speech synthesis. The reported audio-understanding and TTS results use an integration with a Qwen2.5 language-model backbone; they are not standalone capabilities of the tokenizer alone.
- The published speech-reconstruction comparison reports a frame rate of 25 Hz and a bitrate of 325 bits per second for this tokenizer. (github.com)
Availability and limitations
The official repository provides inference and training code, and pretrained checkpoints are available on Hugging Face. The example requires a local software environment and model checkpoints. The tokenizer is a component rather than a complete conversational assistant; downstream task performance depends on the system it is integrated with. Use and redistribution are subject to the project's license terms. (github.com)
Release date
The official README states that the initial GitHub and Hugging Face release was on June 1, 2026. (github.com)
Pricing
As of November 16, 2026, a model-specific hosted API price could not be verified in Tencent Cloud's official mainland China or international pricing materials. Input and output charges, billing units, and CNY or USD prices are therefore unverified. The availability of downloadable code and checkpoints does not establish a hosted API price or eliminate local compute costs. (cloud.tencent.com)