@LonLigrin on Backscroll

2 tweets from @LonLigrin that Backscroll ranked for substance.

Lon Ligrin @LonLigrin 2026-07-08 17:59Z

To reduce whisper hallucinations in noisy/music-heavy audio where VAD fails: Use segment-level no_speech_prob. Try: Drop if >= 0.65, except tiny interjections <= 2 words, <= 0.45s, and < 0.85. Always drop if >= 0.85. No Amara anymore, thank me later

Lon Ligrin @LonLigrin 2026-07-06 17:00Z

WhoSpeaksLive upgrade to diarization pipeline: - Apple Mac support - better clustering - improved ambiguity handling - better speaker reassignment Same accuracy with one embedding provider instead of five. Cleaner evals, less overfitting. https:// youtu.be/uzRdoapMW3c?si =xlpjkbxH5m7Siq5- … via @YouTube

How it ranks: Backscroll reads my Twitter/X timeline and ranks every tweet for substance with an LLM taste rubric — not raw engagement. Curated by Surya Dantuluri.