WhisperX-MLX

WhisperX with MLX acceleration for Apple Silicon - speech recognition with word-level alignment and speaker diarization.

Features

• GPU-accelerated transcription on Apple Silicon using MLX
• Automatic backend selection (MLX on Apple Silicon, faster-whisper elsewhere)
• Word-level timestamps via forced alignment
• Speaker diarization (who said what)
• Voice Activity Detection preprocessing

Installation

# For Apple Silicon (MLX backend)
pip install whisperx-mlx[mlx]

# For CUDA/CPU (faster-whisper backend)
pip install whisperx-mlx[faster-whisper]

# For all backends
pip install whisperx-mlx[all]

Quick Start

Command Line

# Basic transcription
whisperx-mlx audio.mp3 --model large-v3

# With word alignment
whisperx-mlx audio.mp3 --model large-v3 --align

# With speaker diarization
whisperx-mlx audio.mp3 --model large-v3 --diarize --hf-token YOUR_TOKEN

# Force specific backend
whisperx-mlx audio.mp3 --model large-v3 --backend mlx

Python API

from whisperx_mlx import transcribe

# Basic transcription
result = transcribe("audio.mp3", model="large-v3")

for segment in result["segments"]:
    print(f"{segment['start']:.2f} - {segment['end']:.2f}: {segment['text']}")

Models

Supported models:

• tiny, tiny.en
• base, base.en
• small, small.en
• medium, medium.en
• large-v2, large-v3
• turbo (large-v3-turbo)
• distil-large-v3

Requirements

• Python 3.10+
• FFmpeg (for audio loading)
• For MLX backend: macOS with Apple Silicon (M1/M2/M3)
• For diarization: HuggingFace token

License

BSD-2-Clause (same as WhisperX)