cstr/omnivoice-GGUF

OmniVoice GGUF

GGUF conversions of k2-fsa/OmniVoice for the CrispASR omnivoice backend.

Files

File	Size	Description
omnivoice-f16.gguf	1.23 GB	Main model (Qwen3-0.6B LLM + audio embeddings/heads), F16
omnivoice-q8_0.gguf	780 MB	Main model, Q8_0 quantized (embeddings/heads kept at F32)
omnivoice-tokenizer-f16.gguf	403 MB	HiggsAudioV2 audio tokenizer (HuBERT + DAC codec), F16

Usage

# Auto-download
./crispasr --backend omnivoice -m auto --tts "Hello world."

# Manual
./crispasr --backend omnivoice --model omnivoice-q8_0.gguf \
    --codec-model omnivoice-tokenizer-f16.gguf --tts "Hello world."

Status

• Main model GGUF conversion (F16 + Q8_0)
• Qwen3 LLM forward pass (28L, flash_attn)
• Masked iterative code generation (32 steps)
• HiggsAudioV2 DAC decoder (codes to 24 kHz PCM)
• Special token handling (text_start/end, lang_start/end, etc.)
• Audio output: end-to-end text to WAV

Parity note: The C++ generation loop implements the basic masked iterative algorithm. Classifier-free guidance (the unconditional branch that OmniVoice uses for quality) is not yet implemented -- output quality does not yet match the Python reference. The Kaggle parity test confirmed the Python pipeline produces correct speech (ASR roundtrip: exact match).

License

Apache-2.0 (same as k2-fsa/OmniVoice).

Provenance and EU AI Act Art. 53 note

• Upstream model: k2-fsa/OmniVoice — published by k2-fsa.
• Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.