Language auto-detection

Description

Automatically identifying the language spoken in each recording or segment so multilingual research audio is routed to the right transcription settings and analysts.

Whisper predicts the language of the first 30 seconds of audio from 99 languages and returns probabilities. In multilingual research, including code-switching participants, this lets a pipeline route files to language-specific models, translators and analysts, and flag recordings that need human review. Rigour comes from not trusting detection on short or noisy clips, checking detection confidence, detecting language per segment for code-switching speakers rather than per file, and fixing the language explicitly once known, because a wrong detection makes Whisper transcribe or even translate into the wrong language.

Attributes

Loss type
Active
Stage
Discovery
Effort
Hours

Steps

  1. Run detection on a clean speech segment (skip initial silence or moderator introductions) of at least 30 seconds.
  2. Record the top language and its probability; flag files with top probability below about 0.80 for human review.
  3. For known code-switching populations, run detection per segment (e.g. per diarised turn) and store a language tag per segment.
  4. Once a file's language is confirmed, transcribe with the language parameter set explicitly rather than auto-detect.
  5. Report language composition of the corpus and route segments to analysts fluent in each language.

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
  • The Inquiry Series, C9 — Whisper — Batch Transcription at Scale

Also known as

  • Spoken language identification
  • LID

Categories

  • Analytics & Behaviour
  • Computational & Machine Learning