Open streaming memory for real-time voice agents, split into facts and feelings
Emerged finds new products worth a look, daily — what’s new today
Haru
An open 4B speech model that listens and talks at once, in nine languages
Is this yours? Claim itVoice products that need natural interruptions now have an Apache-licensed base model that runs on a 12 GB GPU or an Apple silicon laptop.
Haru-4B is a full-duplex spoken dialogue model from higurashi, a Japanese lab. It hears and speaks in the same 133 ms frames, like a phone call, so turn-taking and overlap come from the model itself. A text card sets the character and a short clip sets the voice. Built on Qwen3.5-4B and trained on over 40 million hours of real speech, it covers nine languages. Code and weights are Apache-2.0, with PyTorch, MLX and vLLM paths. It is a research base model: no benchmarks yet, and it does not follow instructions like an assistant.
More like this
Fast, low-cost speech APIs built on Qwen3 voice models
One open 9B audio model for speech recognition, understanding, TTS and editing
Kokoro's voices with a 7 MB decoder in place of the original 213 MB one
Voice agents that run entirely on CPUs, in the cloud, on-prem or on-device
Video-call an AI character that listens and talks back at once, in your browser