MLX Whisper Pipeline

Transcribe audio files using the MLX Whisper large-v3-turbo model on Apple Silicon.

Features

• Whisper large-v3-turbo — fast, high-quality transcription via mlx-whisper
• Apple Silicon acceleration — native MLX framework (M1/M2/M3/M4)
• Detailed analysis — segment-level metrics and word-level timestamps with probabilities
• Metrics — audio duration, word count, transcription time
• Export — download transcript as plain text or JSON with full segment detail

Requirements

• macOS with Apple Silicon
• Python 3.12+
• FFmpeg
• uv

Setup

brew install ffmpeg
uv sync

Usage

uv run streamlit run streamlit_app.py

Upload or record an audio file (wav, mp3, m4a, ogg, flac, webm, aac) and click Transcribe.

Results appear in two tabs:

• Transcript — full text, copyable
• Detailed Analysis — interactive segment table with per-word detail on row selection

JSON Export

{
  "audio_duration": 10.05,
  "transcript": "transcribed text...",
  "num_words": 42,
  "eval_duration": 3.27,
  "segments": [
    {
      "index": 0,
      "start": 0.0,
      "end": 3.2,
      "text": "transcribed text...",
      "temperature": 0.0,
      "avg_logprob": -0.25,
      "compression_ratio": 1.6,
      "no_speech_prob": 0.01,
      "words": [
        {"word": "transcribed", "start": 0.0, "end": 0.8, "probability": 0.97}
      ]
    }
  ]
}

Duration values are in seconds. eval_duration is rounded to 2 decimal places. audio_duration is null if ffprobe cannot determine it.