This branch contains the unquantized FP32 ONNX quality-baseline runtime.

This repository/branch is a browser/runtime mirror produced by DaisukeDaisuke/tsukuyomichan-omnivoice-onnx.

The converter is open source and its source code, release workflow, verification gates, and conversion scripts are available at: https://github.com/DaisukeDaisuke/tsukuyomichan-omnivoice-onnx

The corrected OmniVoice non-causal attention conversion and the resulting native-Python PoC path were also reviewed and validated with ChatGPT GPT-5.6 Sol high. This is an additional engineering review signal, not a substitute for the repository's reproducible numerical checks, SHA-256 verification, or runtime tests.

The runtime is split into audio_embeddings_encoder, llm_decoder, audio_heads_decoder, and higgs_decoder ONNX graphs with external data. The conversion workflow rejects INT4, INT8, GPTQ, FP16 and BF16 weights/operators and numerically compares exported components against their PyTorch outputs before publication. llm_decoder preserves OmniVoice's rank-4 Boolean non-causal attention mask and runs without KV cache; a 2-D causal/padding-mask LLM contract is rejected by the release gate.

You can run the browser PoC directly here:

The PoC downloads the converted runtime into persistent browser storage so it can be reused across reloads. The current FP32 runtime is about 2.4 GiB, so the first preparation consumes more than 2 GB of network transfer and persistent local storage. Make sure your connection and device have enough capacity before starting the download. Clearing this site's stored data removes the persistent model cache when you want to reclaim the space.

The browser app uses a Service Worker for the offline model-asset path, persistent Cache Storage, and cross-origin isolation. The FP32 validated fast path uses a WebGPU + WebAssembly hybrid: audio embeddings, the LLM, and audio heads run on WebGPU, while the Higgs waveform decoder runs on WebAssembly because this preserves the clean reference audio quality. The LLM FP16 profile keeps the same FP32 graph I/O contract as the FP32 baseline, uses FP16 MatMul weights/compute on the ordinary WebGPU path, and returns each constant MatMul result to FP32 immediately so the remaining LLM math stays FP32. It does not use MatMulNBits. ONNX Runtime Web's JSEP MatMulNBits path does not support the Mobile INT8 LLM on WebGPU, so Mobile INT8 keeps the LLM on WASM while still allowing the unquantized audio embeddings and audio heads to use WebGPU. Browser testing also found reproducible token divergence when the Mobile INT4 LLM alone runs on WebGPU; Mobile INT4 therefore uses the same LLM-on-WASM hybrid while its embeddings and heads remain on WebGPU. Browsers without the required WebGPU path fall back to WebAssembly where supported.

These WAV files are generated on the GitHub Actions CPU runner with native Python + ONNX Runtime from the finalized fp32-unquantized release artifacts. They are static sample files; the player below does not run the model in the browser.

Text: 東京都税関関税許可局、関税許可を急遽却下

FP32:

LLM FP16:

Mobile INT8:

Mobile INT4:

Direct files: FP32 / LLM FP16 / Mobile INT8 / Mobile INT4

Text: WebAssemblyをLLMでVibe Coding中

FP32:

LLM FP16:

Mobile INT8:

Mobile INT4:

Direct files: FP32 / LLM FP16 / Mobile INT8 / Mobile INT4

Text: えへへ、見つけてくれたんだ！ずっとここで待ってたんだよ？

FP32:

LLM FP16:

Mobile INT8:

Mobile INT4:

Direct files: FP32 / LLM FP16 / Mobile INT8 / Mobile INT4

Text: Hey, you finally made it! How does it feel, looking back at everything we've been through?

FP32:

LLM FP16:

Mobile INT8:

Mobile INT4:

Direct files: FP32 / LLM FP16 / Mobile INT8 / Mobile INT4

Use runtime-manifest.json for runtime asset sizes, SHA-256 audit values, and XXH3-128 browser integrity values, and SHA256SUMS for the runtime Release file set. Browser clients intentionally use XXH3-128 for both first-download and reload validation to avoid multi-GiB client-side SHA-256 work, while SHA-256 is retained for CI and offline release auditing. Generated audio samples have their own samples/SAMPLES_SHA256SUMS. The GitHub Release remains the audit/release archive; this Hugging Face repository is the CORS/Range-friendly browser distribution mirror of the same verified files.

本ソフトウェアの音声合成には、フリー素材キャラクター「つくよみちゃん」（© Rei Yumesaki）が無料公開している音声データを使用しています。

このモデルは、つくよみちゃんというキャラクターの声をもとにした音声合成モデルです。 モデルの変換形式や配布形態が変わっても、誰の声をもとにしているのか分からなくならないよう、つくよみちゃん由来であることを明記しています。

■つくよみちゃんコーパス（CV.夢前黎） https://tyc.rei-yumesaki.net/material/corpus/

本モデルから出力した音声は、次の目的ではご利用いただけません。

※鑑賞用の作品として配布・販売していただくことは問題ございません。

出力音声の利用条件の詳細は、つくよみちゃんコーパスの利用規約をご確認ください。 https://tyc.rei-yumesaki.net/material/corpus/

つくよみちゃんのモデルそのものを素材として使用する場合（改変、ファインチューニング、他モデルとのマージ、再配布などを行う場合）、つくよみちゃんコーパスに由来する部分の取り扱いについては、つくよみちゃんコーパスの利用規約に従ってください。この規定は、派生ソフトや再配布されたデータにもコピーレフトされます。

The license: other metadata is intentional. This repository does not claim that conversion to ONNX relicenses the voice checkpoint, Tsukuyomichan corpus-derived materials, upstream OmniVoice model, Higgs Audio 2 materials, or Meta Llama 3 materials. Read TSUKUYOMICHAN_MODEL_CARD.md, TSUKUYOMICHAN_MODEL_CARD_JA.md, OMNIVOICE_MODEL_CARD.md, OMNIVOICE_CODE_LICENSE.txt, BOSON-HIGGS-AUDIO-2-LICENSE.txt, META-LLAMA-3-LICENSE.txt, and NOTICE.txt before use or redistribution.