Posts on X2Q tagged diarization. Updated as new posts are published.
A speaker-identification experiment on phone calls went through five rounds — from a conclusion that 8kHz phone audio 'drowns speaker identity in noise' (EER 35-45%), to a methodology breakthrough (82-92% on a real gold set), to a reality check that crashed to 59.5% on blind calls. The fix wasn't a better model — it was discovering that 68% of 'single-speaker' clips actually contained two voices. Cleaned of those: 87.2% accuracy, AUC 0.856.
pyannote-audio 3.1 works, but requires a Hugging Face token and takes 5 seconds per call. Benchmark of five alternatives on 8 real 2-party phone calls: WeSpeaker + spectral clustering (0.2 s), multi-scale AHC (0.6 s), PLDA + spectral (0.3 s), Silero VAD + spectral (2.1 s) and cross-call gallery nearest-neighbour (0.2 s). The Silero VAD approach won at 76.1% agreement and is 2.4× faster. Surprise: all five collapsed on mono calls — pyannote stays in production.
Business phone calls had been recorded for 13 months — ~33,000 Danish mp3s, ~570 hours of 32 kbps phone audio. The job: transcribe everything, name the speakers, summarize per call/day/week, browsable on a site, and do it locally with no LLM API. This is the build: a benchmark of Danish ASR models, a dual-model + Claude-fusion transcription pipeline, 'phone-first' speaker identification from metadata, self-healing infrastructure across two GPUs, and Claude Code subagents as the (API-less) summary layer. Plus what it teaches about applying the same pipeline to a customer-service function.