Block has introduced Dessa, a full-duplex speech-to-speech model that operates without intermediate text processing. The system is trained exclusively on human speech data, allowing it to process audio inputs and generate audio outputs directly.
The model is capable of listening while speaking, enabling it to manage interruptions and use backchannel cues like 'yeah' or 'got it' during conversations. Demonstrations show the model laughing and adjusting its volume in real-time.
Unlike traditional voice systems that chain speech recognition, text models, and speech synthesis, Dessa handles emotional signals such as sarcasm, laughter, and sighs directly. This approach aims to preserve nuances often lost in text-based pipelines.
The development team notes that training on speech is computationally intensive, as spoken words require significantly more data than their text equivalents. The model is presented as an alternative to existing architectures like Kyutai’s Moshi.
