Richasy/IndexTTS-2.5-GGUF

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.

python tools/convert_index_tts2_5.py `
  --model-dir C:\path\to\IndexTTS-2.5 `
  --output-dir C:\path\to\staging `
  --run-converter C:\path\to\audiocpp_gguf.exe `
  --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:

rin speak hf:Richasy/IndexTTS-2.5-GGUF `
  "这是一段平静而温暖的语音。" `
  --language zh `
  --voice-ref voice.wav `
  --instructions "平静而温暖" `
  --accept-license `
  --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.