Higurashi-AI Releases Haru-4B Full-Duplex Spoken Dialogue Model
Higurashi-AI has released Haru-4B, a 4-billion-parameter multilingual foundation model capable of continuous full-duplex spoken dialogue. It operates at an ultra-low latency of 7.5 Hz, allowing it to listen and speak simultaneously without discrete audio tokens. This open-source release represents a significant technical step forward by replacing traditional turn-based interfaces and separate audio transformers with a single language model. It enables more natural, human-like voice interactions across multiple languages for researchers and developers. Haru-4B processes and generates audio and text frames every 133 ms by summing them into a single position of the language model. It utilizes MeanFlow sampling through a small MLP head to produce a 64-dimensional latent audio representation in a single sampling step, supporting nine different languages.
## BACKGROUND
Traditional spoken dialogue systems typically operate in a half-duplex, turn-based manner where the system must wait for the user to finish speaking before generating a response. Full-duplex conversational systems allow for simultaneous listening and speaking, closely mimicking natural human conversation dynamics.