aufklarer/Kyutai-STT-1B-MLX-8bit

Kyutai STT 1B — MLX INT8

Native MLX export of Kyutai's delayed-stream speech recognition model. Derived from kyutai/stt-1b-en_fr; attribution: Kyutai.

Property	Value
Model size	1B class
Decoder precision	INT8; embeddings and output head FP16
Format	MLX safetensors, speech_swift_v1
Sample rate / frame size	24 kHz mono / 1,920 samples
Attention window / trained text delay	750 frames / 0.5 s
Audio tokenizer	Mimi, 32 codebooks, original unquantized checkpoint
Runtime bundle size	1.483 GiB

Files

File	Size	Description
README.md	4.55 KiB	This model card
SHA256SUMS	0.64 KiB	SHA256 checksums of the other eight files
VALIDATION.md	20.09 KiB	Methods, regression checks and limitations
benchmark.json	267.26 KiB	Per-utterance quality, latency and memory measurements
config.json	1.53 KiB	Architecture, timing, precision and provenance
export.json	0.29 KiB	Pinned source and export record
mimi.safetensors	366.83 MiB	Original Mimi codec weights
model.safetensors	1151.78 MiB	STT decoder weights
tokenizer_en_fr_audio_8000.model	117.56 KiB	SentencePiece tokenizer

Performance and validation

Measurement	Result
FLEURS English / French	80 / 80 test clips
English word error rate (lower is better)	23.74%
French word error rate (lower is better)	17.13%
MLX free-buffer cache limit	256 MiB
Machine / build	Apple M5 Pro, 48 GB, macOS 26.6.2; Swift release, compiled Metal resources

See benchmark.json for every transcript and timing, and VALIDATION.md for methods, regression checks and limitations. Quiet-input residual codebook decisions and complete transcripts can differ from the original PyTorch codec. Corpus WAV files preserve the original Float32 samples. The corpus measurements use sequential batches of 10 clips, with a fresh process and warmup for each batch. These corpus sweeps check quality; speed and memory figures below come from the separate isolated CLI profile.

CLI cache-policy profile

The corpus table above uses a 256 MiB MLX reusable-buffer cache. The Kyutai CLI defaults to a 256 MiB reusable-buffer cache. This separate profile uses the first eight English and eight French clips (157.7 s). These timing and memory measurements have a different sample count from the corpus table. This profile uses PCM16 copies of the recordings; its WER is separate from the source-preserving Float32 corpus results.

Measurement	Result
MLX free-buffer cache limit	256 MiB (Kyutai CLI default)
Profile word error rate	15.72%
Real-time factor, including final flush	0.130
Throughput	7.72× real time
Peak resident memory	1.418 GiB
Peak physical footprint	1.906 GiB
Median first-text compute (unpaced input)	179.3 ms
Median input consumed at first text	1.6 s
Median of per-clip frame p95 (80 ms audio frames)	9.1 ms

The limit controls cached buffers, not total process memory. The CLI restores the previous process-wide policy on return; the model API leaves it to the caller.

The runtime supports file transcription and incremental PCM input. Semantic VAD extra heads are not included.

Usage

With a speech-swift build that includes the KyutaiSTT module:

speech transcribe recording.wav --engine kyutai --model aufklarer/Kyutai-STT-1B-MLX-8bit
speech transcribe recording.wav --engine kyutai --model aufklarer/Kyutai-STT-1B-MLX-8bit --stream
speech transcribe recording.wav --engine kyutai --model aufklarer/Kyutai-STT-1B-MLX-8bit --mlx-cache-limit-mb 256

From Python, invoke the same local runtime:

import subprocess
subprocess.run(["speech", "transcribe", "recording.wav", "--engine", "kyutai",
                "--model", "aufklarer/Kyutai-STT-1B-MLX-8bit"], check=True)

Source and links

Weights: Kyutai, CC-BY-4.0. Implementation reference: Moshi and delayed-streams-modeling. Native runtime: speech-swift. Documentation · soniqo.audio · blog.