Description
Automatically segmenting an interview recording by speaker, so transcripts distinguish participant from moderator and analysis uses only participant speech.
Speaker diarisation answers 'who spoke when'. pyannote.audio uses neural voice activity detection, speaker change detection and speaker embeddings clustered into speakers, producing time-stamped speaker turns that can be aligned with Whisper's word timestamps. In UX research this separates moderator prompts from participant answers, supports focus groups and co-interviews, and prevents moderator language from contaminating coding and NLP. Rigour comes from specifying the expected number of speakers when known, measuring diarisation error rate on a reference sample, and handling overlapping speech, which remains the main error source. Speaker labels are anonymous (SPEAKER_00); mapping them to roles must be checked.
Attributes
- Discipline
- Computational & Machine Learning
- Loss type
- Active
- Goal
- Synthesise data
- Stage
- Discovery, Evaluation
- Data
- Qualitative
- Effort
- Hours
Steps
- Prepare 16 kHz mono WAV files and accept the model licence terms; run locally for sensitive data.
- Run the pyannote speaker-diarisation pipeline, passing num_speakers when known (e.g. 2 for one-to-one interviews).
- Align speaker turns with Whisper word timestamps to produce speaker-attributed transcripts.
- Map anonymous speaker labels to roles (moderator, participant) using a rule such as who speaks first, and verify on every file.
- Measure diarisation error rate on 5-10 hand-labelled files and investigate files with heavy overlap.
- Exclude or flag moderator turns before coding and NLP analysis.
Sources
- whisper.cpp (local, privacy-preserving)
- openai-whisper (cloud)
- pyannote.audio (diarisation)
- ffmpeg
- pydub
- Bredin, H. et al. (2020) pyannote.audio: neural building blocks for speaker diarization. ICASSP
- Radford, A. et al. (2023) Robust speech recognition via large-scale weak supervision. ICML
- The Inquiry Series, C9 — Whisper — Batch Transcription at Scale
Also known as
- Who-spoke-when segmentation
- Speaker segmentation and clustering
Categories
- Analytics & Behaviour
- Computational & Machine Learning