---

license: cc-by-4.0

language:

• en
• fr

pipeline_tag: automatic-speech-recognition

tags:

• audio
• speech-recognition
• transcription
• gguf
• kyutai
• moshi
• mimi
• codec

library_name: ggml

base_model: kyutai/stt-1b-en_fr

---

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

Provenance and EU AI Act Art. 53 note

• Upstream model: kyutai/stt-1b-en_fr — published by kyutai.
• Upstream licence: cc-by-4.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (GGUF/GGML). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.