xiry.

AI · AI model

Dessa

Full-duplex speech-to-speech AI model trained entirely on human conversation without text.

From the Xiry library · Updated 24 October 2026

Visit block.xyz ↗

Dessa is a conversational AI model developed by Block that was trained from scratch strictly on human speech rather than text. Because it operates audio-in and audio-out in full-duplex mode, it can listen and speak simultaneously with realistic timing, backchanneling, and emotional tone. The model runs faster than real time and is compact enough to execute locally on high-end Apple silicon hardware.

What it's for

  • Build natural voice-first conversational interfaces without text-to-speech latency
  • Experiment with speech models that handle natural interruptions and emotional inflection
  • Run real-time conversational voice agents locally on high-spec Apple hardware

Features

  • Full-duplex audio processing that listens and speaks simultaneously
  • Trained strictly on raw sound without text transcripts or pretrained models
  • Natural conversational dynamics including interruptions, laughing mid-sentence, and backchanneling
  • Runs locally on a MacBook Pro with M4 Max or better
  • Faster-than-real-time inference latency
  • Trained on a single NVIDIA GB200 NVL72 setup

Pricing

Free. Currently in research preview / waitlist; no pricing announced.

Details

Type
AI model
Category
AI
Topic
Speech Models
Open source
No
Platforms
macOS, Web
For
AI researchers, voice interface designers, and developers building conversational audio apps.
Alternative to
OpenAI Advanced Voice Mode, Kyutai Moshi, Hume AI
Tags
speech, voice, audio, full duplex, speech to speech, conversational ai

Why it was saved

  • @jack, 23 October 2026: Real, locally runnable frontier speech model for building voice-first products and natural conversational interfaces.

Related

Updated