Block’s Dessa Learns Conversation From Speech Alone

Block has introduced Dessa, an experimental conversation model that the company says was trained from scratch on human speech rather than text or pretrained models. Its central idea is to preserve the timing and expression that can disappear when a voice assistant converts speech into a transcript. This is a research demonstration, not a generally available assistant.

Why avoid the transcript?

A common voice-assistant pipeline converts audio into words, asks a language model for a response, then synthesizes speech. That design is useful, but a transcript is a compressed account of a conversation: a long pause, an interruption or a strained voice may carry meaning that the words alone do not capture. Not every existing voice system uses this same pipeline.

Block describes Dessa as audio-in, audio-out, with conversational knowledge learned from speech. Its examples include laughter, changing loudness and emotionally expressive responses. TechCrunch independently reported the announcement and described the examples, but that reporting is not an independent benchmark of their reliability.

What full-duplex changes

Full-duplex means that listening and speaking can overlap. Instead of waiting for a rigid turn to end, an assistant can theoretically register a correction while responding, or give a brief acknowledgment while someone continues. The important distinction is interaction timing, not simply how realistic a synthesized voice sounds.

The official page presents recordings as live conversations and says subtitles were added afterward. These selected demonstrations help illustrate the intended behavior. They do not establish how often the model interrupts badly, misses a correction or fails in a noisy room.

Local does not mean publicly available

Block says inference can run locally on a MacBook Pro with an M4 Max chip or better and faster than real time. That is a hardware and speed claim from the developer, not a published cross-device performance evaluation. It is not a promise that every consumer computer can run the model.

Local inference can reduce the need to send conversational audio to a server, depending on the actual application. It does not by itself establish an application’s privacy policy, recording behavior or data retention. Block’s page offers a waitlist and early access for a small group of safety researchers, rather than unrestricted public access.

What still needs testing

Block also says training used one NVIDIA GB200 NVL72 system and hundreds of times fewer words than large language models. Those figures are company-reported and do not provide an apples-to-apples comparison of capability, data quality, compute cost or energy. A single system is not the same thing as a single GPU.

The useful next tests are less theatrical than a moving demo: response latency across devices, performance during interruptions, accents and background noise, factual accuracy, and whether users can tell they are speaking to a machine. A voice that sounds caring is not evidence of feelings, understanding or suitability for clinical advice.

For now, treat Dessa as evidence of a promising research direction, not a ready replacement for a trusted service. If a product eventually ships, check availability, hardware requirements and recording controls before adopting it.