tts-rs-assets

Derived assets for tts-rs: per-stage reference activations that the Rust ports are gated against, plus two small CosyVoice artifacts that cannot be produced without the upstream python package.

No model weights are here. You download those from their own repositories; see docs/setup.md in tts-rs.

What is in here, and why it is not in the git repo

file	size	what it is
cosyvoice/rand_noise.safetensors	4.6 MB	The CFM decoder's fixed noise tensor. CausalConditionalCFM builds it once under set_all_random_seed(0) and slices the same tensor every call, so it is not in the checkpoint — a port that samples its own noise looks correct and sounds different.
cosyvoice/tokenizer.json	10.6 MB	A consolidated HF tokenizer. The checkpoint's added_tokens_decoder lists three specials; CosyVoice3Tokenizer.__init__ registers ~250 more at construction (`<
fixtures/audio8/oracle.safetensors	65 MB	Per-stage fp32 activations dumped from Audio8's PyTorch reference.
fixtures/audio8/default_voice_codes.safetensors	20 KB	Reference speech codes for the default voice.
fixtures/cosyvoice/oracle.safetensors	41 MB	Per-stage fp32 activations dumped from CosyVoice's PyTorch reference.
fixtures/qwen3tts/oracle.safetensors	10 MB	Per-stage fp32 activations dumped from Qwen3-TTS's PyTorch reference.

The first two exist because producing them requires the upstream CosyVoice repository on PYTHONPATH under python 3.10 / torch 2.3.1. Hosting them means running CosyVoice needs neither. The fixtures exist so that ./scripts/gates.sh — the thing that substantiates every correctness claim in the tts-rs README — runs on a machine with no PyTorch at all.

Both remain reproducible from source: references/cosyvoice/convert.py and each engine's dump_fixtures.py regenerate them. See docs/setup.md §3. If you are checking the ports rather than using them, regenerating is the stronger move — verifying against tensors somebody else uploaded is trusting the upload.

Integrity

SHA256SUMS covers every file. scripts/fetch-assets.sh in tts-rs verifies against it and refuses a mismatch.

shasum -a 256 -c SHA256SUMS

Provenance

Dumped from these upstream models. The Rust ports were written against these revisions; a fixture mismatch after an upstream update is a revision difference, not necessarily a port bug.

model	revision	licence
Audio8/Audio8-TTS-Preview-0.6b	f9612f13a0ab40facf3d050fc908b9e6db05c2be	Apache-2.0
FunAudioLLM/Fun-CosyVoice3-0.5B	29e01c4e8d000f4bcd70751be16fa94bf3d85a18	Apache-2.0
Qwen/Qwen3-TTS-12Hz-1.7B-Base	fd4b254389122332181a7c3db7f27e918eec64e3	Apache-2.0

Those SHAs are each repository's HEAD as read on 2026-08-12, recorded so that future drift is diagnosable. The dumps themselves were taken in late July / early August 2026; if an upstream repository moved between the dump and that reading, the SHA above is the later one. Regenerate from source if you need certainty.

Licence

Apache-2.0, inherited from all three upstream models. Every file here is a derivative work of an Apache-2.0 model — activations computed by it, a tokenizer built by its code, or a tensor drawn by its constructor. Attribution is above; see NOTICE in the tts-rs repository.