IndexTTS-1.5 — GGUF (ggml-quantised)

GGUF / ggml conversion of IndexTeam/IndexTTS-1.5 for use with CrispStrobe/CrispASR.

IndexTTS-1.5 is a zero-shot voice cloning TTS system: reference audio + text → cloned speech at 24 kHz. Architecture: Conformer conditioning encoder (6L, d=512) → Perceiver resampler (2L, 32 latents) → GPT-2 AR decoder (24L, d=1280, 20 heads) → BigVGAN vocoder (6-stage upsample, anti-aliased SnakeBeta activations). ~500M parameters total. Distributed under Apache-2.0 license.

Two GGUF files are needed: the GPT model (conditioning + text → mel codes → latent) and the BigVGAN vocoder (latent → 24 kHz audio).

Files

All quant levels produce correct speech (ASR roundtrip = "Hello world!"). F16 gives 100% mel-code parity with Python; Q8_0 is the best quality/size trade-off.

Quick start

# Easiest: auto-download (~870 MB on first run)
./build/bin/crispasr --backend indextts -m auto \
    --voice reference_speaker.wav \
    --tts "Hello world, this is IndexTTS speaking."

# Output: tts_output.wav (24 kHz mono)

Or with explicit paths:

# 1. Build CrispASR
git clone https://github.com/CrispStrobe/CrispASR
cd CrispASR
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j --target crispasr-cli

# 2. Pull model files (pick your preferred quant)
huggingface-cli download cstr/indextts-1.5-GGUF indextts-gpt-q8_0.gguf --local-dir .
huggingface-cli download cstr/indextts-1.5-GGUF indextts-bigvgan.gguf --local-dir .

# 3. Synthesise with voice cloning
./build/bin/crispasr --backend indextts \
    -m indextts-gpt-q8_0.gguf \
    --codec-model indextts-bigvgan.gguf \
    --voice reference_speaker.wav \
    --tts "Hello world, this is IndexTTS speaking."

Features

• Zero-shot voice cloning — any 3-10 second reference WAV at 24 kHz (or auto-resampled)
• Multilingual — trained on English and Chinese; cross-language cloning works
• Beam search — num_beams=3 with repetition_penalty=10.0 (matches Python defaults)
• No external dependencies — SentencePiece tokenizer embedded in GGUF, no espeak/phonemizer needed

Accuracy

With the F16 model, C++ mel codes match Python 100% (55/55 tokens identical to the Python greedy reference). Conditioning norm matches Python within 0.001%. All quantization levels (Q8_0, Q4_K) produce correct speech verified via ASR roundtrip.

Architecture details

Input text → uppercase → SentencePiece unigram tokenizer (12000 vocab)
Reference audio → 24kHz resample → mel spectrogram (100 bands, hop=256)
  → Conformer encoder (6 blocks, d=512, 8 heads, Conv2d subsampling)
  → Perceiver resampler (2 layers, 32 latents, d=1280, GEGLU FFN)
  → 32 conditioning vectors

GPT-2 AR decoder:
  [32 cond latents | text_embs + text_pos | start_mel + mel_pos]
  → 24 transformer blocks (d=1280, 20 heads, GELU FFN)
  → gpt.ln_f → final_norm → mel_head → beam search (B=3)
  → mel codes (stop token = 8193)

Latent extraction (2nd pass):
  Full sequence → GPT-2 → gpt.ln_f → final_norm → [n_mel+1, 1280]

BigVGAN vocoder:
  latent [T, 1280] → conv_pre → 6× (ConvTranspose1d + AMPBlock1 with
  anti-aliased SnakeBeta) → conv_post → tanh → 24kHz PCM
  + ECAPA-TDNN speaker embedding for voice conditioning

Conversion

python models/convert-indextts-to-gguf.py \
    --model-dir /path/to/IndexTTS-1.5 \
    --output indextts-gpt.gguf \
    --vocoder-output indextts-bigvgan.gguf

License

Apache-2.0 (same as upstream IndexTTS).

Citation

@misc{indextts2024,
  title={IndexTTS: An Industrial-Level Zero-Shot Text-to-Speech System with Controllable Timbre},
  author={IndexTeam},
  year={2024},
  url={https://github.com/index-tts/IndexTTS}
}