Speech-to-speech
Audio goes in and audio comes out without forcing a spoken moment through text first.
Independent research guide + realtime demo
Dessa AI Voice Model
A guide to Block's speech-to-speech research preview, plus an independent realtime voice playground for exploring timing, tone, interruption, and natural conversation.
This site is independent and not affiliated with Block. The interactive demo is powered by Qwen Realtime, not the unreleased Dessa model. Read the official research at block.xyz/dessa.
Audio in / out
No text bottleneck
Fully duplex
Listens while speaking
Emotion-aware
Keeps vocal signals
Local research
M4 Max or better
Voice, kept in its original medium
Explore the behaviors that make speech-native AI feel less like a transcript and more like a conversation.
Fully duplex
Dessa can listen while responding, enabling overlap and natural backchannels.
Emotion signals
Cadence, breath, hesitation, and emphasis remain available to the model.
Natural turn-taking
Short acknowledgements and pauses help a voice interaction find its rhythm.
Human-trained
The research preview learned from human conversation rather than a text-only corpus.
Safety-aware research
A practical guide to the limits, access model, and open questions around voice AI.
Why the architecture matters
A conversation is more than its transcript
Traditional assistants split a conversation into several separate systems. Speech-native research asks what becomes possible when audio remains available throughout the interaction.
DimensionSTT → LLM → TTSRealtime audio model
Signal pathThree separate stagesOne continuous session
TimingWaits for a turn to endResponds during live audio
InterruptionOften restarts a turnSupports natural barge-in
Voice signalReduced to transcriptAudio context stays available
How this demo works
A transparent Qwen-powered path
This independent product demonstrates realtime voice interaction with Qwen. It explains Dessa research, but it does not claim to run Block's unreleased model.
01
Browser audio
Your microphone is captured as mono PCM and played back as streaming audio.
02
Secure gateway
A short-lived signed session keeps the provider key out of the browser.
03
Qwen Realtime
The model listens, responds, transcribes, and supports interruption in one session.
Browser supportCurrent Chrome, Edge, and Safari with microphone permission.
Data handlingTranscript and usage metadata are saved; audio storage is not enabled.
Session limitsUp to 15 minutes, further limited by your available credits.
Realtime voice plans that scale with you
Start with welcome credits, then choose a transparent voice-minute plan for the Qwen-powered realtime demo.
1 credit covers up to 1 started minute. Sessions are capped by your available balance and saved to your private history. These plans do not sell access to Block's Dessa research model.
Free
$0
Explore the product before you pay
- Welcome credits after sign-up
- No card required
- Private transcript history
Starter
$9/mo
60 realtime voice minutes each month
- 60 voice minutes / month
- Up to 15 minutes per session
- Private transcript history
- Email support
Popular
Pro
$29/mo
300 realtime voice minutes for regular use
- 300 voice minutes / month
- Up to 15 minutes per session
- Private transcript history
- Copy and download transcripts
Team
$79/mo
1,200 minutes for growing teams
- 1,200 voice minutes / month
- Private transcript history
- Copy and download transcripts
- Priority support
Questions around the research
A concise reference before you explore the demo or the technical notes.
Hear what a text-free voice model changes.
Try the browser demo, then explore the independent guides for builders, researchers, and curious listeners.