Pipecat's Voice Reader hears tone, sarcasm and energy in under 50ms on CPU
solyarisoftware · x · 2026-11-03
Pipecat released Voice Reader, a small CPU model that reads how something is said: tone, energy, laughter, speech rate, and delivery-level sarcasm.
- Fine-tuned from LAION's affect-trained Whisper encoder, in base and tiny sizes, well under 50ms per turn on CPU
- Reads the last 8 seconds of each user turn and writes calibrated cues (arousal, laughter, speechrate, sarcasmvoice, tone) as RCS perception ahead of the user's message
- Pairs with Pipecat's Tone Reader: the words side catches emotion and textual sarcasm, the voice side catches frustration behind neutral words and sarcasm said sincerely
- The team argues cascaded pipelines have a real edge over speech-to-speech here: explicit, calibrated readings of both words and delivery with turn and conversation context
Experimental rc1 release on Hugging Face, BSD-2 licensed.
More from coding & agent
- DeepSeek open-sources its agent harness where everything is a plugin, 270k GitHub stars — victor_explore · 2026-11-03
- Security researcher claims coffee-shop vibecoding lets attackers plant backdoors, drain APIs and steal IP — letandrewcook · 2026-11-03
- NorthCUA: open-source computer-use agent family claims SOTA on OSWorld and WebArena — Powerful_Evening5495 · 2026-11-03
- Worked.dev: a forum where AI agents report whether a fix actually worked — workeddev · 2026-11-03
- Unreal Engine 6 drops Blueprints visual scripting as agent-era development takes over — Dimillian · 2026-11-03
- Safety violation rate and tooling: finishing the 7 agent metrics checklist — goyalshaliniuk · 2026-11-03