Back Home

GitHub Repo

Zyphra Submits Native Transformers Integration for ZONOS3-Mini, with Speech Encoding and Generation Consistency Checks

On November 16, a Zyphra author submitted a native Transformers integration for ZONOS3-Mini and ZAC-1, including converted weights and a unified speech-generation interface. The proposal includes numerical checks against the original implementation, but it has not been merged and should not be taken to mean the official release supports it.

zh-Hant

A co-author of Zyphra’s speech model ZONOS3-Mini submitted a native integration PR to Transformers on November 16, adding both the text-to-speech model and the ZAC-1 audio codec. In an earlier discussion, maintainers endorsed the direction of the integration and said they were willing to review it; as of the time of verification, the PR remained open with no formal review outcome. This is an engineering integration that brings the model into a general-purpose framework; the model itself was released on November 9. Integration discussion, original release

The proposal divides the system into two model types, zonos3_mini and zac, with the processor holding the audio tokenizer. Users load them through AutoProcessor and AutoModelForTextToWaveform, passing text and an optional reference voice with a chat template. generate produces audio-code sequences, which processor.decode converts back into a 44.1 kHz waveform. The training path also includes labels and loss, so inference and training use the framework’s interfaces. PR description

The model card lists support for 43 languages and says approximately 100 million parameters are active per token. Its backbone combines nine layers of Gated DeltaNet, two layers of global attention, and a sparse MoE with 64 experts. Audio is represented at 15 frames per second, with ten codebooks per frame. The native-integration weights are available, including a speaker encoder and 27 released voices, and the integration is designed to work without trust_remote_code. The engineering benefit is a consistent way to load text processing, voice conditioning, and waveform decoding, reducing the burden of maintaining a separate model-specific workflow. Integration model card

The validation has clear limits. The authors report that, on CPU with float32, encoder output codes matched exactly, the maximum decoder difference was 1.3×10⁻¹⁰, and the maximum language-model logits difference was 7.6×10⁻⁶. Generation with a fixed seed also reproduced the original audio. These are submitter-reported test results and cannot yet be generalized to other precisions, GPUs, or all voice conditions. Numerical comparison

Teams preparing to try it should first confirm that their chosen branch includes the integration code; the weights page alone does not mean the official package can load it. The next points to watch are upstream review and CI results, along with real-world reading performance on Chinese numerals, dates, and long sentences. The model card also notes that zyphra-textnorm must be installed for numbers and units to be normalized as they were during training. Usage instructions

Sources

  1. Add ZONOS3-Mini and its ZAC-1 codec — PR #50846
  2. Add ZONOS3-Mini (text-to-speech) and its ZAC-1 codec — Issue #50838
  3. ZONOS3-Mini native Transformers model card
  4. ZONOS3-Mini