Tyger: Open-Weights Streaming ASR and Full-Duplex Turn-Control Voice Model
Developers have released Tyger, an open-weights model combining NVIDIA Nemotron 3.5 ASR Streaming with a causal GRU to handle real-time, full-duplex turn-control every 80 ms. Alongside the model, they introduced the tyger-s2s stack, which integrates Gemma 4 E4B FP8 and VoxCPM2 to power local voice agents with barge-in handling and tool-calling capabilities. This release tackles a fundamental challenge in conversational AI by allowing voice agents to listen and speak simultaneously without confusing their own playback for user input. It lowers the barrier for building responsive, low-latency local voice assistants capable of natural interruptions. The Tyger model operates every 80 ms without lookahead to choose between answering, stopping, or chiming in, while tracking both microphone and playback audio to prevent echoes. However, the complete tyger-s2s stack requires approximately 28 GB of VRAM, is currently restricted to FP32, and has only been benchmarked offline on an RTX 5090.
## BACKGROUND
Full-duplex voice agents are systems capable of listening to a user while simultaneously generating and playing back speech, enabling more natural conversational dynamics like interruptions. Streaming ASR (Automatic Speech Recognition) allows speech to be transcribed incrementally in real time as audio streams in, rather than waiting for an utterance to finish.