Block’s new speech system, Dessa, runs on consumer‑grade GPUs and can speak and listen at the same time while reproducing cues like laughter and crying for a more natural AI conversation.

Block has unveiled Dessa, a full‑duplex voice model that can simultaneously speak and listen while mirroring human emotional cues such as laughter, sighs, and even crying, promising a more natural conversational AI experience on consumer‑grade GPUs.

What Makes Dessa Different

Unlike traditional text‑to‑speech systems that operate in a single direction, Dessa’s architecture supports real‑time bidirectional audio flow, allowing it to react to a speaker’s tone and adjust its own output on the fly. The model is trained on a diverse dataset of emotional speech, enabling it to embed subtle affective signals that make interactions feel less robotic.

Technical Highlights

  • Runs on off‑the‑shelf GPUs, removing the need for specialized hardware
  • Full‑duplex processing reduces latency to under 200 ms
  • Supports a range of emotional expressions, from joy to melancholy

The system leverages a transformer‑based backbone optimized for low‑precision inference, which keeps power consumption comparable to standard video‑game graphics workloads while delivering high‑fidelity audio.

Potential Applications

Developers can integrate Dessa into virtual assistants, interactive games, and remote‑learning platforms where nuanced vocal feedback enhances user engagement. Its ability to convey empathy could also benefit mental‑health chatbots that need to respond sensitively to user emotions.

For readers interested in broader AI trends, the AI coverage on uxc.news provides context on how voice technologies are reshaping human‑computer interaction.

Dessa feels like talking to a person who truly understands your mood, not just a scripted response.

Block’s rollout includes an open API, inviting third‑party developers to experiment with the model’s expressive capabilities while maintaining strict privacy safeguards for user data.

Read the full story on TechCrunch coverage of Block’s new voice model.

Track coverage on Momentum · More in AI.