OmniVoice-MLX-fp16
A genuine fp16 MLX conversion of k2-fsa/OmniVoice —
a massively multilingual (600+ languages) zero-shot TTS built on a single-stage
NAR discrete-diffusion architecture (Qwen3 backbone run bidirectionally) with a
Higgs-audio v2 codec.
Converted for the Soniqo speech stack (speech-swift / MLX). ~3.3 GB fp32 → ~1.5 GB fp16.
Contents
| File | What |
|---|---|
model.safetensors | Qwen3 backbone (llm.*) + audio embeddings + 8 codebook heads |
audio_tokenizer/model.safetensors | Higgs-audio v2 codec (semantic wav2vec2 + acoustic DAC RVQ) |
tokenizer.json, config.json, chat_template.jinja | Qwen3 text tokenizer + config |
Float tensors are cast to fp16; the PyTorch key structure is preserved.
Validation
The fp32 source model round-trips at ~0% WER (whisper-base.en) on a 5-sentence English set; this bundle is a faithful half-precision copy.