Downloads · 30 days
12
18% of all-time downloads
jiangzhuo9357/omnivoice-onnx-bidi-bf16
omnivoice-onnx-bidi-bf16 is a text-to-speech model from jiangzhuo9357. Use it when you need text read aloud. It is set up for onnxruntime. The card lists the license as cc-by-nc-4.0.
Self-contained bf16 variant of the corrected bidirectional ONNX export of k2-fsa/OmniVoice (600+ language zero-shot voice-cloning TTS, Qwen3-0.6B backbone, 32-step non-autoregressive iterative unmasking, Higgs Audio V…
Downloads · 30 days
12
18% of all-time downloads
All-time downloads
68
Public
Repo size
2 GB
Likes
0
Public
Click a slice to open those files.
.onnx1.9 GB · 98%
From the Hugging Face model README
Self-contained bf16 variant of the corrected bidirectional ONNX export of
k2-fsa/OmniVoice (600+ language zero-shot voice-cloning TTS,
Qwen3-0.6B backbone, 32-step non-autoregressive iterative unmasking, Higgs Audio V2 codec, 24 kHz). Produced for
Sokuji (issue #351).
Sibling variants: -bf16 (default; fastest + clean
on CUDA), -fp32 (un-quantized reference),
-int4 (smallest).
bf16 (default, recommended) — fastest AND clean on the onnxruntime CUDA EP. A naive fp16 export emits an all-zero llm_decoder output on CUDA (deep-layer RMSNorm x^2 overflows fp16's 65504 range → inf → rsqrt 0 → zeros); bf16's fp32-range exponent has no such overflow. audio_embeddings is also bf16 (verified lossless, embeds cosine 1.000).
llm_decoder.onnx, audio_embeddings_encoder.onnx, audio_heads_decoder.onnx + tokenizer/config — the backbone.audio_tokenizer/ — Higgs Audio V2 codec graphs (fp32).voices/ — curated reference clips (.wav + transcript) for built-in cloning voices.Runs on plain onnxruntime (no onnxruntime-genai); the non-autoregressive decode loop lives in the app.
OmniVoice is non-autoregressive: its 32-step iterative unmasking uses the full (bidirectional) context. The
community onnx-community/OmniVoice-Onnx export bakes in a causal mask (via onnxruntime-genai's
Qwen3ForCausalLM) → every MASK position collapses to one token → noise, not speech. This repo re-exports the
llm_decoder with the correct bidirectional mask (hidden-state cosine ≈ 1.000 vs the fp32 reference).
License: CC-BY-NC-4.0. Inherited from k2-fsa/OmniVoice (weights are
CC-BY-NC due to the Emilia training set, CC-BY-NC-4.0). This constraint originates upstream and cannot be relaxed
by any re-upload or format conversion. This is an unofficial re-export for interoperability — not affiliated
with nor endorsed by the OmniVoice authors. Some intermediate re-uploads mislabel OmniVoice apache-2.0; that is
incorrect for the weights — do not use commercially.