Disclosure: I'm Chaozi, and this is an independent write-up. I'm not affiliated with Block.
Block has been publishing research on speech-to-speech models: systems that take spoken audio in and produce spoken audio out, without chaining separate speech-to-text, language model and text-to-speech stages. That direct approach can cut latency and keep tone, pacing and emphasis that a text pipeline throws away.
I put together Dessa AI Guide as an independent reference for people who want a plain-language look at this model. It covers what speech-to-speech means, why it matters for voice assistants and real-time conversation, and what to watch as the technology matures.
Why it matters
- Lower latency: fewer hand-offs between components means faster replies.
- More natural speech: prosody and emotion can carry through.
- Simpler systems: one model instead of a three-stage pipeline.
Read more
If you're curious, visit Dessa AI Guide. It's an unofficial community resource, so please check Block's own announcements for authoritative details.
Feedback is welcome in the comments.
Top comments (2)
The latency reduction of true speech-to-speech is a massive leap. Traditional STT → LLM → TTS pipelines always destroy the emotional nuance and pacing, plus that extra 500ms of latency makes conversations feel like walkie-talkies.
Since you put this guide together, did you look into how Block's model handles "barge-in" (when a user interrupts mid-sentence)? That's usually where S2S models completely fall apart compared to human conversation.
speech-to-speech from Block, curious how it holds up outside demos.