cstr/f5-tts-GGUF

F5-TTS v1 Base — GGUF

Native C++ GGUF conversion of SWivid/F5-TTS (MIT license) for the CrispASR runtime.

Model file

File	Size	Description
f5-tts-v1-base-f16.gguf	953 MB	F16 weights for DiT/Vocos, F32 for critical paths (AdaLN, RoPE, time embed)

Note on quantization: F5-TTS uses a 32-step iterative ODE solver where each step runs the full 22-layer DiT twice (for CFG). This means every weight matrix is used 1408 times per synthesis. Q8_0's ~0.5% per-operation error compounds multiplicatively across these passes, producing unintelligible output — even when the conditioning pathway (AdaLN, timestep MLP) is kept at F32. F16's ~0.001% error survives the 1408× accumulation. This is inherent to flow-matching architecture, not a ggml limitation. The converter supports --quant q8_0 for experimentation, but F16 is the only recommended format.

Architecture

• DiT backbone: 22-layer Diffusion Transformer with AdaLN-Zero (330M params)
• Text encoder: Character-level ConvNeXtV2 (4 blocks, 512-d)
• Vocoder: Vocos (8× ConvNeXt + ISTFTHead, 13M params)
• ODE solver: 32-step Euler with CFG (strength=2.0, sway=-1.0)
• Output: 24 kHz mono PCM
• Voice cloning: Zero-shot from 3-15s reference audio + transcript

Single GGUF contains both DiT and Vocos — no separate codec model needed.

Usage

# Install / build CrispASR
git clone https://github.com/CrispStrobe/CrispASR && cd CrispASR
cmake -B build && cmake --build build -j$(nproc) --target crispasr-cli

# Synthesize with voice cloning
./build/bin/crispasr --backend f5-tts -m auto \
    --voice reference.wav \
    --ref-text "Transcript of the reference audio" \
    --tts "Hello, how are you today?" \
    --tts-output output.wav --seed 42

The --ref-text flag is required — F5-TTS conditions on both audio and its transcript for voice cloning.

Conversion

Converted from SWivid/F5-TTS safetensors using:

python models/convert-f5-tts-to-gguf.py \
    --model-dir /path/to/f5-tts \
    --output f5-tts-v1-base-f16.gguf

Quantization beyond F16 is not recommended for flow-matching models.

License

MIT (same as upstream SWivid/F5-TTS).