Local Whisper deployment (whisper.cpp)

Description

Running the Whisper speech recognition model on a researcher's own machine so interview recordings are transcribed accurately without sending participant audio to a cloud service.

Whisper is a transformer speech recognition model trained on about 680,000 hours of multilingual audio, robust to accents and background noise. whisper.cpp is an efficient C/C++ implementation that runs on laptops and workstations (CPU, Apple Silicon or GPU) using quantised model weights. Local deployment keeps raw participant audio on controlled hardware, which simplifies consent, GDPR and data processing agreements and is often required in health, finance and public-sector research. Rigour comes from pinning the model size and version (e.g. large-v3 for accuracy, medium or small for speed), setting language and decoding parameters explicitly, recording them for provenance, and auditing a sample of transcripts against human references, since Whisper can hallucinate fluent text during silence or noise.

Attributes

Loss type
Active
Effort
Hours

Steps

  1. Check consent and data protection terms, and confirm that local processing satisfies them.
  2. Build whisper.cpp and download a pinned model (e.g. ggml-large-v3, or a quantised q5 variant for constrained hardware), recording its checksum.
  3. Convert audio to 16 kHz mono WAV, which whisper.cpp expects.
  4. Run transcription with explicit language, timestamps and output formats (txt, srt or json), and enable voice activity detection where available to reduce hallucination in silences.
  5. Batch-process recordings with a script that logs model, parameters, runtime and any errors per file.
  6. Audit a stratified sample of transcripts (e.g. 10% or at least 5 files) for word error rate before analysis.

Sources

  • whisper.cpp (local, privacy-preserving)
  • openai-whisper (cloud)
  • pyannote.audio (diarisation)
  • ffmpeg
  • pydub
  • Radford, A. et al. (2023) Robust speech recognition via large-scale weak supervision. ICML
  • Gerganov, G. whisper.cpp (GitHub repository)
  • The Inquiry Series, C9 — Whisper — Batch Transcription at Scale

Also known as

  • On-device speech-to-text
  • Local automatic speech recognition

Categories

  • Analytics & Behaviour
  • Computational & Machine Learning