IndexTTS 2.5 original-dtype GGUF
This repository contains one original-dtype standalone GGUF converted from the official
IndexTeam/IndexTTS-2.5 snapshot for use with
audio.cpp release 0.6 and Rodel.Inference.
It is a derivative work, not an official IndexTeam or bilibili release. Read the governing
Chinese license, its
English translation,
NOTICE, and
third-party notices before downloading or using it.
Artifact
| File | Bytes | SHA-256 |
|---|
IndexTTS2.5-GGUF/index-tts2_5-orig.gguf | 7,885,093,568 | 07e9bfe77bd42b6e67b8e5a39b365635f7198951281990cb49f322ef546cb9b4 |
No quantized package is published here. The first release deliberately keeps every tensor in its
original dtype so quality can be measured before adding another approximation.
Reproducible conversion
Sources:
- model:
IndexTeam/IndexTTS-2.5@c39ce5ba981572cb187443877ff559dfb246ce63
- converter:
0xShug0/audio.cpp@bb15edd78b56e035967e0eb999a6b28a62337db4
- Wav2Vec2-BERT:
facebook/w2v-bert-2.0@da985ba0987f70aaeb84a80f2851cfac8c697a7b
- CAMPPlus:
funasr/campplus@e4b6ede7ce16997aff4ae69fbca1f0175e2afede
- BigVGAN:
nvidia/bigvgan_v2_22khz_80band_256x@633ff708ed5b74903e86ff1298cf4a98e921c513
The source LFS files were checked against the SHA-256 values published by Hugging Face. The
converter was built from the release source with MSVC 19.51, /utf-8, CPU-only GGML, and only the
index_tts2 model target.
1python tools/convert_index_tts2_5.py `
2 --model-dir C:\path\to\IndexTTS-2.5 `
3 --output-dir C:\path\to\staging `
4 --run-converter C:\path\to\audiocpp_gguf.exe `
5 --type orig
audiocpp_gguf --inspect reports 3,790 tensors, all ten required namespaces, embedded sidecars,
embedded model spec, and family index_tts2. A raw tensor comparison against audio.cpp's
original-dtype package found 3,790 equal tensors and zero differences. The whole-file SHA-256 is
different because unordered JSON object keys in the embedded model spec were serialized in a
different order by the Windows build; tensor content is identical.
The machine-readable conversion and governance contract is
rin-audio-model.json.
Rodel.Inference
The first download requires explicit acceptance of the attached model license:
1rin speak hf:Richasy/IndexTTS-2.5-GGUF `
2 "这是一段平静而温暖的语音。" `
3 --language zh `
4 --voice-ref voice.wav `
5 --instructions "平静而温暖" `
6 --accept-license `
7 --play
The model requires a reference WAV and an explicit zh, en, ja, es, or ar language. It
does not consume a reference transcript. Instructions are mapped to emotion text.
Known limitations
- audio.cpp does not fully port the official WeText and Japanese normalization pipeline. Numeric
expressions, percentages, fractions and dates require preprocessing; Rodel refuses numeric
notation instead of producing a known-wrong reading.
- audio.cpp PR 210 records an unresolved Japanese repetition-penalty parity question.
- The official IndexTTS tracker contains reports of weaker timbre, emotion and prosody in 2.5 than
in 2.0. A structurally correct GGUF is not itself a quality claim.
- Streaming audio, duration control, emotion reference audio and eight-dimensional emotion vectors
are not part of the initial Rodel integration.
Measured Rodel validation
The package was acquired from the public Hub through Rodel's immutable tree and digest path, then
run with the verified audio.cpp 0.6 CUDA profile + runtime bundle:
- 22.05 kHz mono PCM16 WAV, with no native worker left after private or daemon shutdown;
- Chinese sample: 2.71 seconds generated in 1.89 seconds, RTF 0.699;
- warm attached request: RTF 0.452;
- 6,300 English characters: 265.01 seconds generated in 52.60 seconds, RTF 0.198;
- independent Whisper large-v3-turbo transcription: exact zh/en/es, 0.947 normalized-character
similarity for ar and 0.909 for ja.
These measurements prove the package, explicit language routing and intelligibility, not voice
identity or emotion quality. Those still require speaker-similarity measurement and human
listening.
License
This repository is governed by the attached bilibili Model Use License. LICENSE_ZH.txt is the
governing Chinese text; LICENSE is the English translation. Distribution must retain both, the
derivative disclaimer, and the files under THIRD_PARTY_LICENSES/, and must bind downstream
recipients to the same terms. Separate authorization is required above 100 million monthly active
users or RMB 100 million annual revenue under the Chinese terms.
