Ex-OpenAI engineer: telephony vendors' 8kHz audio silently degrades voice models
pbbakkum · x · 2026-11-13
pbbakkum highlights a subtle issue most evals miss: voice models perform worse with less audio information. gpt-live-1 was specifically trained for 8kHz performance to mitigate this, but lower fidelity still hurts. He calls out popular telephony vendors that only offer 8kHz instead of 16kHz, arguing the industry should just fix this and move on.
Related event: Check SIP trunk audio is 16kHz before feeding voice models(3 posts)→
More from Models
- Microsoft and Claude go down simultaneously, taking apps offline with them — ThePeterMick · 2026-11-13
- Insider teases: Google still has a 'big juicy' model release up its sleeve — iruletheworldmo · 2026-11-13
- 1B model tops MathArena under-12B leaderboard with 83.62 — tensorqt · 2026-11-13
- Daily AI brief: Gemini 4 Hydrogen pricing revealed, Anthropic to list as ANT on Nasdaq Nov 19 — testingcatalog · 2026-11-13
- Unknown Darwin-180B-RSI posts perfect 100 on MathArena HMMT leaderboard, raising eyebrows — tensorqt · 2026-11-13
- After two months daily-driving Claude Code, one reviewer says Opus 5.5 beats GPT everywhere except computer use — kimmonismus · 2026-11-13