Block's Dessa is also unique, a large-scale model trained entirely on human speech. It's a true audio-in-audio-out model, with no text conversion throughout, trained from scratch using only human conversation recordings (no pre-trained model). It achieves full-duplex (speaking while listening), natural timing, emotion, interjections, and backchannel (e.g., "yeah," "got it"). This differs from most existing voice AI: OpenAI/Google typically follow a cascaded path of speech → text → LLM → speech, losing prosody and emotion; even Kyutai's Moshi started with a text language model and predicted text in parallel. It pushes voice AI from a cascaded architecture to native audio models, potentially improving the naturalness and emotion perception of customer service and assistants (security-related), and proving that small teams with relatively limited computing power can explore new paradigms. This also suggests that there is still much unexplored space for different modalities and training paths.