Reverb-TTS (ONNX export)
Unofficial ONNX export of the transformer of software-mansion/Reverb-TTS
(297M masked-diffusion TTS derived from k2-fsa/OmniVoice, Qwen3 layers),
so it runs in ONNX Runtime (CPU or Vulkan/WebGPU) without PyTorch. Weights are exported from the
fp16 checkpoint and stored as fp32 graphs. License: CC BY-NC 4.0 (non-commercial, attribution), as the original.
Not affiliated with Software Mansion. See NOTICE.
The audio codec is not in this repo: it is the Higgs Audio v2 tokenizer, exported separately in melser/Higgs-Audio-2-Tokenizer-onnx under its own license.
Files (onnx/)
| Graph | Inputs -> outputs | Size |
|---|---|---|
core.onnx |
x[1,S,1024], past_k/past_v[24,1,6,Pp,128] (Pp may be 0), bias[S,Pp+S] -> hidden, new_k, new_v |
1.06 GB |
embed_prefix.onnx |
text_ids[1,Pt], ref_codes[1,8,R] -> x[1,P,1024] |
98 MB |
embed_target.onnx |
tokens[1,8,F] -> x[1,F,1024] |
34 MB |
heads.onnx |
hidden[1,F,1024] -> log_probs[1,8,F,1024] (fp32, mask id excluded) |
34 MB |
SHA256SUMS lists checksums. Tokenizer, preset voices (voices/*.json) and the Spanish/English prompt format come
from the original repo; the sampler is reverb_tts/sampler.py (pip install reverb-tts).
How the graphs are used
- Prompt
<|denoise|><|lang_start|>{lang}<|lang_end|><|instruct_start|>None<|instruct_end|><|text_start|>{reference_text} {text}<|text_end|>-> token ids; reference audio -> 8-codebook codes with the codec encoder. - Prefix:
embed_prefix->core(empty past, zerobias) -> keepnew_k/new_v. - 8 main + 2 polish passes over the fully masked
[8,F]target:embed_target->core(past = prefix K/V;bias0, orpolish_bias=1.0 on the reference columns in polish passes) ->heads; unmask/remask persampler.py. - Codes -> waveform with the codec decoder (24 kHz).
Verification (CPU, 2026-11-20)
- Driving the official
reverb_tts.sampler.sample()through these graphs gives 100 % identical codes (all 8 codebooks, two prompts) and a bit-identical waveform versus the PyTorch path. - AMD BC-250 (Vulkan/RADV, ORT WebGPU EP):
core/headsmatch CPU (argmax 100 %), 4-6x faster. Keep the codec decoder on CPU there (it returns near-silence on that GPU). - Spanish word error rate (Whisper small, 4 texts x 6 seeds): 0.000 with presets
es_male_standard_1/2, 0.031 cloned from a clean 7 s clip.
Conversion scripts and parity checks are in scripts/ (requirements.txt pins the versions used).
Model tree for melser/Reverb-TTS-onnx
Base model
Qwen/Qwen3-0.6B-Base