Transcribe interview recordings locally and review suspicious passages
At a glance
A researcher published a local Whisper pipeline with transcript exports, optional name masking and review artifacts. This workflow is listed under Local Transcription, Interview Research, Document Workflows.
- Agents / tools
- Whisper, Python, secure_local_HPC_speech_transcription
SourceBuilding a tool for these workflows? See weekly rates.
What they did
1) Use Python 3.11 or newer and a local copy of pablobernabeu/secure_local_HPC_speech_transcription. Create and activate a virtual environment, install requirements.txt and run python -m preflight to check model dependencies and audio decoding. A GPU is optional for the local path. 2) Place a recording at audio_input/interview.wav. Run python transcription.py audio_input/interview.wav --language english --timestamps --mask-personal-names --redacted-timestamps-only, changing the language to match your recording. Models must be available locally for an offline run. 3) Inspect the text/Word transcript and the redacted timestamp exports under output/transcripts. Read the generated review JSON and listen again to flagged passages before accepting the transcription. 4) Check the masking against the recording and keep the manifest with the outputs. Later, use python -m provenance verify followed by the manifest path to check whether exported artifacts changed.
Worth knowing
Name masking misses names outside its dictionary; review flags are heuristics. Speaker attribution additionally needs a Hugging Face token and model download.