Sparrow-2
A turn-taking model that knows when a voice agent should speak
Emerged finds new products worth a look, daily — what’s new today
Rising on Hacker News· 17 points
If your voice agent feels sluggish between turns, this cuts roughly a tenth of a second off every reply in exchange for a few more false stops.
RealtimeVAD keeps Silero for deciding whether someone is speaking, then adds a quicker test every 10 ms that compares the audio level with the room's noise floor and the speaker's own level. When the voice drops away, it can call the end of speech before Silero's probability falls. It emits start, pause and end events for 16 kHz mono audio, needs only numpy and onnxruntime, and bundles the Silero model so torch is not required. MIT, installed with pip, and still alpha at 0.0.5.
More like this
Voxel
An open-source desktop voice changer that feeds any app through a virtual mic
JARVIS
A voice assistant that brainstorms with you, then runs and watches Claude Code
CleanTake
Patch a ruined line of speech with the same moment from your other recorder
Nari Qwen3-TTS and Qwen3-ASR
Fast, low-cost speech APIs built on Qwen3 voice models
Egma
Open-source simulation testing and monitoring for voice agents