THE SHORT ANSWER
● Available today — The TTS pipeline. Dessa is restricted to AI safety researchers
● Keeps tone and timing — Dessa, because it never converts audio to text
● Talks and listens at once — Dessa - full-duplex by design. Pipelines take turns
● Runs on your laptop — Dessa, on a MacBook Pro M4 Max or better
Every voice product in production today runs the same three stages: speech to text, then a language model, then text to speech. Block announced Dessa on 23 October with none of those stages. It takes audio in and returns audio out, and it was trained from scratch on recorded human conversation with no text and no pretrained models anywhere in the process.
The two architectures
| TTS pipeline (current) | Dessa (audio-native) |
|---|
| Stages | Three - STT, LLM, TTS | One |
| Training data | Text corpora plus paired audio | Conversation audio only |
| Turn-taking | Sequential - waits for a turn to end | Full-duplex - listens and speaks at once |
| What survives the trip | Words. Tone, overlap and timing are discarded at transcription | The sound itself |
| Training hardware | Large clusters | One NVIDIA GB200 NVL72 |
| Local inference | Rarely practical end to end | MacBook Pro M4 Max or better |
| Availability | Generally available, priced | AI safety researchers, institutional email |
| Pricing | Published by most vendors | None announced |
What transcription throws away
This is the substance of Block's claim. When a pipeline transcribes speech, it keeps the words and discards everything the words were carried in: hesitation, emphasis, overlap, how fast someone is speaking, whether they are about to stop. The language model then reasons over a transcript that has already lost the information a human listener uses most. Dessa's argument is that if you never transcribe, you never lose it.
Block also says Dessa learned from hundreds of times fewer words than a large language model, which follows from the same design - conversation audio is a much denser signal per word than text.
Full-duplex is the part that is immediately testable
Sequential pipelines have to decide when you have finished talking, which is why voice assistants interrupt you or sit silently after you stop. A full-duplex model does not make that decision because it is listening while it speaks. Anyone who has used a current voice assistant will recognise the failure mode this removes.
Why this is not yet a purchase
Dessa is behind an application gate for AI safety researchers with institutional email, with a public waitlist and launch links described only as coming soon. No pricing, no latency figures, no benchmark against the pipelines it is arguing with. The twelve-person lab behind it reports directly to Jack Dorsey, which tells you about commitment rather than timeline.
For anything shipping this quarter, a published TTS stack remains the answer. Dessa is the thing to watch for next year.