Overview
Whisper transcribes speech to text with near-human accuracy across accents, noise, and 100 languages — and because the weights are open, it runs entirely on your machine. For interview-based research, it converted a per-hour-billed service industry into a local batch job with zero confidentiality exposure.
Why researchers use it
- Local processing — interview audio never leaves your machine (ethics boards approve)
- Multilingual — transcription and English translation in one model
- Free at any scale — a hundred interview hours costs electricity
Getting started
- For speed use faster-whisper (pip install faster-whisper) or whisper.cpp on Macs.
- Transcribe: a few lines of Python or one CLI call per audio file; batch a folder overnight.
- Add WhisperX or pyannote for speaker diarisation (‘who said what’) — the missing piece for interviews.
The honest review
Strengths. Accuracy that holds up on real fieldwork audio, timestamps for quote-finding, and the compliance story: your consent forms’ ‘data stays local’ clause remains true.
Limitations. Raw output needs cleanup for QDA (fillers, punctuation, speaker turns); diarisation requires the add-on tools; GPU strongly recommended for large models at scale.
When NOT to use this
Always human-verify quotes before publication — models mishear names and negations, exactly the words that change meanings. Transcription is solved; verified transcription is still your job.