From conversation to conversation
Most voice products coordinate three components: automatic speech recognition, a text language model, and speech synthesis. That system is practical and measurable, but every conversion can discard information or add delay.
Block says its Dessa AI model was trained directly on human conversation. In principle, an audio-native representation can retain not just the words but also how they were spoken, giving the model more context for timing and response style.
Why full duplex matters
A full-duplex system can listen while it speaks. That enables backchannels such as “yeah” or “got it,” immediate interruption when the user changes direction, and turn-taking that does not depend on a long silence.
Duplex behavior is not only a model feature. Browser audio capture, echo cancellation, streaming transport, voice activity detection, and playback scheduling all contribute to whether the experience feels natural. A strong AI model can still feel slow when the surrounding system is poorly designed.
Training efficiency and local inference
Block has reported that Dessa learned language from substantially fewer words than large text models and was trained on a single Nvidia GB200 NVL72 system. It has also stated that local inference can run on a MacBook Pro M4 Max or better.
Those statements are technically interesting but should remain attributed to Block until model weights, evaluation methods, or broader researcher results are available. Hardware support, sustained latency, memory use, and safety behavior are all important areas for independent verification.
Safety questions for a speech-native AI model
Voice contains signals that text systems can miss, including distress, pain, urgency, and uncertainty. Better sensitivity may improve safety, but it also increases the importance of consent, retention policies, access controls, and clear limits on what the system can infer.
A responsible product should disclose which model is running, where audio is sent, what is stored, how usage is billed, and how users can delete their history. This site exposes those distinctions rather than presenting the Qwen-powered demo as Dessa.