Beyond Transcription: Building Voice AI That Understands Conversations
summary
Shows concrete gaps and fixes: running Nvidia Parakeet on AMI headset reports 11.4% WER, the same model on the table mic yields 26% WER on AMI, and pyannoteAI’s Precision‑2 hits 3% DER versus 5% baseline. Demos include a two‑person phone call with overlapping words, a reconciliation pipeline called STT orchestration that interleaves speaker‑attributed words, and an exclusive diarization heuristic to pick which speaker STT should transcribe. Repository and notebooks are available for repro.
AI-written summary · about
chapters (7)
- 0:00Introduction to pyannote
- 2:03Beyond basic transcription
- 7:50Understanding diarization
- 12:27Evaluating diarization models
- 16:07Attributed transcription
- 19:43STT orchestration demo
- 22:43Conclusion and Q&A
related talks
16mThe End of Awkward AI Transcriptions
08mQ n A with Dr Ananth Sankar
36mNVIDIA NeMo: Toolkit for Conversational AI
31m“Sorry, Could you repeat that again?” - Speech Recognition with Python
20mVoice fingerprinting & recognition
22mBuilding next-gen speech synthesis for Bookmate audiobook service
47mThe Digital Hearth: An In-the-Wild Automated Group Speech to Text System
43mTalk by Om Thakkar (OpenAI)