Local Batch Transcription Pipeline (Whisper)

Description

Transcribing large interview sets locally with Whisper, adding speaker diarisation and auditing word error rate, keeping audio within your environment.

Transcription is a bottleneck and a privacy risk when audio is uploaded to third-party services. A local pipeline runs speech recognition on the researcher's own machine or controlled server. Audio is normalised (for example 16 kHz WAV via ffmpeg), transcribed in batch with whisper.cpp or Whisper with explicit language settings and word-level timestamps, and segmented by speaker using pyannote.audio diarisation aligned to the timestamps. Quality is audited by sampling transcripts, computing word error rate against human correction, and retranscribing or correcting where WER exceeds a threshold such as 10 percent. Documenting that audio never left the environment supports ethics approvals and data protection commitments.

Attributes

Loss type
Active
Effort
Hours

Steps

  1. Local setup: Install whisper.cpp locally (does not send audio to external servers). Convert audio to WAV at 16kHz: ffmpeg -i input.mp3 -ar 16000 output.wav
  2. Batch transcription: Run whisper.cpp on all files. Set language explicitly. Request word-level timestamps.
  3. Speaker diarisation: Run pyannote.audio to identify speaker turns. Align with Whisper timestamps to produce speaker-labelled transcripts.
  4. Quality audit: Sample 3 transcripts per 20 for human review. Calculate Word Error Rate (WER). If WER > 10%: retranscribe manually.
  5. Privacy documentation: Running entirely locally = no audio leaves your environment. Document this in ethics reporting.

Sources

  • whisper.cpp (OpenAI)
  • pyannote.audio
  • ffmpeg
  • Robust Speech Recognition via Large-Scale Weak Supervision (Radford et al.)

Also known as

  • Whisper transcription
  • Automated speech recognition pipeline
  • whisper.cpp with diarisation

Categories

  • Analytics & Behaviour
  • Computational & Machine Learning