cstr/kyutai-stt-1b-GGUF

Kyutai STT 1B (en/fr) -- GGUF

GGUF conversions and quantisations of kyutai/stt-1b-en_fr for use with CrispStrobe/CrispASR.

Available variants

File	Quant	Size	Notes
kyutai-stt-1b.gguf	F16	2.0 GB	Full precision
kyutai-stt-1b-q8_0.gguf	Q8_0	1.1 GB	High quality
kyutai-stt-1b-q4_k.gguf	Q4_K	636 MB	Best size/quality tradeoff

All variants produce correct transcription on test audio (JFK speech).

Model details

• Architecture: Mimi neural audio codec encoder (SEANet CNN + 8-layer transformer + RVQ with 32 codebooks at 12.5 Hz) + 16-layer causal transformer LM (2048d, RoPE, SwiGLU, RMSNorm)
• Parameters: ~1B
• Languages: English, French
• Audio input: 24 kHz mono (auto-resampled from 16 kHz)
• License: MIT
• Reference: moshi.cpp (MIT)

This is a novel codec-based ASR architecture: audio is first encoded into discrete tokens via the Mimi neural audio codec, then a causal language model autoregressively predicts text tokens from the audio codes. Unlike encoder-decoder models (Whisper, Parakeet), the entire pipeline is autoregressive.

Usage with CrispASR

git clone https://github.com/CrispStrobe/CrispASR && cd CrispASR
cmake -S . -B build && cmake --build build -j8

# Auto-detect backend from GGUF
./build/bin/crispasr -m kyutai-stt-1b-q4_k.gguf -f audio.wav

# Explicit backend
./build/bin/crispasr --backend kyutai-stt -m kyutai-stt-1b-q4_k.gguf -f audio.wav -osrt

Conversion

python models/convert-kyutai-stt-to-gguf.py --input kyutai/stt-1b-en_fr --output kyutai-stt-1b.gguf
crispasr-quantize kyutai-stt-1b.gguf kyutai-stt-1b-q4_k.gguf q4_k