A 1.6B causal Transformer audio tokenizer for speech, sound, and music.
Unified discrete audio tokenization with semantic alignment and streaming support.
MOSS-Audio-Tokenizer is a 1.6-billion-parameter causal Transformer audio tokenizer designed for next-generation audio foundation models. It suits researchers and audio engineers building advanced text-to-speech, speech recognition, or generative audio systems.
It compresses raw audio into a low frame rate of 12.5Hz using a pure Transformer architecture built entirely from causal blocks without CNNs. Trained on 3 million hours of diverse data, it jointly optimizes encoder, decoder, RVQ, discriminator, and an LLM for semantic alignment in a single pipeline.
It suits machine learning researchers and audio developers building custom speech or music generation pipelines. It does not suit casual users looking for a ready-to-use desktop audio editor or simple plugin.
Built entirely from causal Transformer blocks without CNNs to support low-latency streaming.
Employs a 32-layer Residual Vector Quantizer supporting bitrates from 0.125kbps to 4kbps.
Produces discrete tokens that achieve both lossless-quality reconstruction and high-level semantic alignment.
Provides PyTorch, ONNX Runtime, and TensorRT inference backends for flexible deployment.
Free under its open-source license, with the only cost being the hardware infrastructure required to host and run a 1.6B parameter model.
897 open remote AI-training roles, updated November 17, 2026; most pay $20–$170 an hour. Zekai lists them free; the application goes to the employer’s platform.
These reviewed AI tools do the same job instantly — from our scored directory.
It was trained on 3 million hours of diverse audio data covering speech, sound effects, and music.
Yes, the repository provides ONNX Runtime and TensorRT inference backends for deployment without PyTorch.
See Zekai first in Google