Block announces Dessa, a full-duplex conversational model trained entirely on human speech Other

Block has announced Dessa, an end-to-end full-duplex conversational speech model trained entirely on recordings of human conversation without text tokens or pretrained foundation models. Developed by a specialized research team reporting directly to Block Chairman Jack Dorsey, the model listens and speaks simultaneously, capturing emotional tone, conversational interruptions, and natural laughter.

Most current conversational voice assistants rely on an automatic speech recognition (ASR) step that transcribes speech into text before handing it to a language model. This conversion discards critical audio cues such as inflection, sarcasm, emotional distress, and hesitation. Dessa bypasses text entirely by operating natively on audio waveforms as both input and output.

The model was trained from scratch using a single NVIDIA GB200 NVL72 rack. According to Block, Dessa acquired conversational competence from a speech dataset containing hundreds of times fewer words than typical text LLMs consume during pretraining. Inference executes faster than real time and runs locally on Apple Silicon systems equipped with an M4 Max chip or higher. Block has opened an early-access waitlist focused on AI safety researchers before expanding availability.

What it is

• Speech-native pretraining. Trained from scratch exclusively on human speech recordings without relying on text corpora or base text models.
• Full-duplex conversation. Listens and speaks concurrently, allowing mid-sentence interruptions, conversational backchanneling ("yeah", "got it"), and real-time laughter.
• Emotion and prosody preservation. Natively interprets and generates sarcasm, vocal inflection, laughter, and emotional nuances lost during text transcription.
• Single-rack training footprint. Pretrained on a single NVIDIA GB200 NVL72 rack with high data efficiency compared to text token pipelines.
• Local Apple Silicon inference. Operates faster than real time locally on macOS hardware with Apple M4 Max processors or better.

Technical specifications

Sources

• Block: Dessa Conversation Model Research Preview