beshkenadze/mms-lid-256-mlx

MMS-LID-256 — MLX

Meta's MMS-LID-256 language identification model running on MLX (Apple Silicon Metal GPU).

Identifies 256 languages from raw audio. Uses weights directly from facebook/mms-lid-256 — no conversion needed.

Performance (M1, 10s audio)

Framework	Latency	Russian	English
Python MLX (with mx.compile())	267ms	98.8%	99.8%
Swift MLX (with compile())	268ms	97.3%	99.7%
CoreML GPU	250ms	89.1%	99.8%

Usage

# Clone the benchmark repo
git clone https://github.com/beshkenadze/lid-bench
cd lid-bench/mlx

# Setup
uv venv && uv pip install mlx numpy soundfile safetensors huggingface_hub

# Download weights (from original facebook repo)
huggingface-cli download facebook/mms-lid-256 --include "model.safetensors" "config.json"

# Run
python mms_lid_256.py path/to/audio.wav --benchmark

Full implementation: github.com/beshkenadze/lid-bench

Model Details

• Architecture: Wav2Vec2ForSequenceClassification (48 transformer layers, 16 heads)
• Input: Raw 16kHz waveform (zero-mean unit-variance normalized)
• Output: 256 language probabilities
• Parameters: 315M
• Weight format: Original HF safetensors (no conversion needed)
• Weight loading: Conv1d axis swap + weight_norm precomputation done at load time

Notes

• Weights are loaded directly from facebook/mms-lid-256 HF cache — no separate conversion step
• This repo contains only the model card and MLX implementation reference
• ANE causes 13x slowdown — use Metal GPU (.cpuAndGPU) only
• Sustained inference on M1 degrades to ~400ms due to thermal throttling (48 transformer layers). M2+ should be better.