Problem: An organisation needed accurate transcripts of meetings with speaker identification, sometimes with several conversations in the room. Off-the-shelf services misheard the technical vocabulary and could not separate speakers reliably.

Approach: We deployed an end-to-end ASR pipeline on NVIDIA NeMo with speaker diarisation. Audio preprocessing handles noise reduction and format conversion from whatever the source was. A web frontend serves searchable, timestamped transcripts.

Technical details:

• NVIDIA NeMo ASR with language model adaptation for technical vocabulary
• Speaker diarisation with voice profile clustering
• Preprocessing: noise reduction, normalisation, multiple input formats
• Web frontend with search, filtering by speaker, and export
• Configurable backend for other ASR engines

Outcome: Past meetings are searchable by speaker and topic, so nothing depends on somebody's notes. Accuracy on technical vocabulary beats the off-the-shelf services it replaced.

Stack: Python, NVIDIA NeMo, ASR, Diarisation, Web Frontend

Sandboxed plugin ran for projects/speech-to-text and read “Speech-to-Text Pipeline” through the mediated API.