AI news · October 18, 2026
Tyger lets voice agents listen while they speak
Tyger's open-weight voice agent handles interruption and backchannels on one GPU; test a short booking flow before using it for real calls.
- creator-reported reply time
- 250 ms
- GPU for the core stack
- 1
NineNineSix presented Tyger as a full-duplex voice agent that keeps listening while it talks. Its demo shows the agent stopping when a person cuts in, waiting through a pause, saying a short acknowledgement without taking the turn, and calling a function. The site says replies arrive in about 250 milliseconds and offers both an API route and open weights with commercial use included.
The linked Tyger-S2S repository splits the system into EARS and MOUTH on one GPU. EARS handles speech and turn decisions; MOUTH uses Gemma 4 E4B and VoxCPM2 to answer and speak while the reply is being made. The project is creator-reported and still asks for a Hugging Face token, so the useful first test is a short local conversation with clear success and stop rules, not a real customer call.
What you can do with it
Run a five-turn booking script with one interruption, one pause, one correction, and one function call. Score stop timing, missed words, duplicate speech, function accuracy, response time, and manual recovery. Keep the full-duplex mode only if it reduces correction work without creating false actions.
Our take
Full-duplex speech can make an agent feel much more natural because it handles the messy parts of a conversation. It also makes wrong turns harder to notice. Start with read-only or reversible functions, keep a visible transcript, and require confirmation before any action with a real consequence.
Links Try itRun the code
Source: NineNineSix — Made With Models writes the brief; the reporting is theirs.