Measure a real turn
Time microphone stop to first audible answer, then test interruption. A demo can feel responsive while failing under network loss or a long prompt.
Research-preview guide · checked Oct 27, 2026Dessa AI Voice:
Dessa AI Voice:
what is real today?
Dessa has sparked interest because it frames voice interaction as a live conversation, not a chain of transcription, text generation, and playback. This guide separates the public research story from tools a developer can actually ship with today.
Dessa is a research story first—not a production API to build a business on yet.
Block’s public Dessa material describes a speech-first, full-duplex AI research effort. As of this page’s review date, do not assume there are public model weights, a self-serve API, commercial pricing, or permission to market another product as “Dessa.” Check the original announcement for the current access status before planning a build.
What “full duplex” means
A full-duplex voice system can listen while it speaks and handle interruption, overlap, timing, and vocal cues. It is a product interaction pattern, not a guarantee that a system is factually accurate, safe for impersonation, or ready for every production use case.
If you need a voice product now, choose for the workflow—not the headline.
| Need | Evaluate first | What to validate before launch |
|---|---|---|
| Live spoken assistant | OpenAI Realtime API or another documented real-time platform | Interruption behavior, latency, transcript storage, disclosure, and escalation to a human. |
| Script-to-voice narration | ElevenLabs documentation or another licensed TTS provider | Voice rights, commercial-use terms, output controls, and per-character cost. |
| Multimodal assistant in Google’s stack | Gemini Live API documentation | Regional availability, modality limits, safety controls, and pricing on the exact model. |
Disclose synthetic voices
Get consent for recording, do not clone a person without rights, and make it clear when a caller is speaking to AI—especially in support or financial contexts.
Price the whole conversation
Include input audio, output audio, model reasoning, retries, storage, and human handoff. “Per minute” is rarely the only cost.
Three claims that should remain separate.
“It sounds human”
That is a subjective reaction to samples. It does not establish reliability, factual accuracy, or suitability for a specific customer workflow.
“It runs locally”
A research demonstration does not mean the public can download weights, use a commercial license, or run it on ordinary hardware.
“It can replace support staff”
That needs measured task success, audit logs, consent, recovery paths, and quality review—not only a natural-sounding voice.
Common Dessa AI questions
Can I download Dessa AI or use a public API?
Do not assume so. Access, weights, pricing, and API availability can change. Use Block’s original page as the source of truth rather than a third-party site offering an unverified download.
Is Dessa an alternative to text-to-speech?
Not exactly. Text-to-speech reads prepared text. A full-duplex voice agent also needs turn-taking, listening, response generation, safety controls, and a product workflow around the voice.
Can I make a website around the trend?
Yes—an independent explainer, access tracker, and alternatives guide can be useful. Do not imply affiliation with Block, sell access you do not have, or use the name as your product brand.
Sources and reading
- Block: Dessa — primary source to re-check for current access and product status.
- OpenAI Realtime API documentation.
- Google Gemini Live API documentation.
Compare the cost behind a voice product.
Before choosing a model, estimate text-model usage and understand the cost components you control.