One neuron per language hits 99.1% language-ID accuracy, beating dedicated models
julianweisser · x · 2026-11-20
New neuron interpretability work on speech models (Phonon-2.5) surfaces striking findings:
- A model trained only to transcribe contains a language detector: a single neuron per language reaches 99.1% identification accuracy, beating dedicated language-ID models (97.2%) — while probing a whole encoder block yields just 27%
- Other neurons act as a built-in speedometer tracking speech rate, and one can sometimes infer what kind of room you're in from audio alone
- The team traced why the model "hears" words when nobody is speaking and fixed it by editing just 10 KB of weights — roughly 50× fewer parameters than a traditional VAD
More from Models
- Naval-war game benchmark: author claims GPT 6.2 Astra crushes Opus 5.2 — RichardsonDx · 2026-11-20
- Qwen3.8-27B-pi-adapted GGUF quantization trends on Hugging Face — harness-hub · 2026-11-20
- Google's Genie 4 goes live on Gemini API: steerable 720p worlds via WebRTC — _philschmid · 2026-11-20
- 10 hands-on lessons on what actually happens inside an AI after you type a prompt — techartist_ · 2026-11-20
- Redwood questions Google reusing safety results from weaker model for stronger Carbon — YvesMulkers · 2026-11-20
- GTE-MLM sparse embedding models now natively supported in Transformers v5 — tomaarsen · 2026-11-20