Whisper Transcribe

Transcribe audio files to text using OpenAI Whisper. Supports speech-to-text with auto language detection, multiple output formats (txt, srt, vtt, json), batch processing, and model selection (tiny to large). Use when transcribing audio recordings, podcasts, voice messages, lectures, meetings, or any audio/video file to text. Handles mp3, wav, m4a, ogg, flac, webm, opus, aac formats.

Transcribe audio with scripts/transcribe.sh:

# Basic (auto-detect language, base model)
scripts/transcribe.sh recording.mp3

# German, small model, SRT subtitles
scripts/transcribe.sh --model small --language de --format srt lecture.wav

# Batch process, all formats
scripts/transcribe.sh --format all --output-dir ./transcripts/ *.mp3

# Word-level timestamps
scripts/transcribe.sh --timestamps interview.m4a

Models

Model	RAM	Speed	Accuracy	Best for
tiny	~1GB	⚡⚡⚡	★★	Quick drafts, known language
base	~1GB	⚡⚡	★★★	General use (default)
small	~2GB	⚡	★★★★	Good accuracy
medium	~5GB	🐢	★★★★★	High accuracy
large	~10GB	🐌	★★★★★	Best accuracy (slow on Pi)

Output Formats

• txt — Plain text transcript
• srt — SubRip subtitles (for video)
• vtt — WebVTT subtitles
• json — Detailed JSON with timestamps and confidence
• all — Generate all formats at once

Requirements

• whisper CLI (pip install openai-whisper)
• ffmpeg (for audio decoding)
• First run downloads the model (~150MB for base)