Block, the payments and fintech company led by Jack Dorsey, has introduced Dessa, an artificial intelligence voice model the company says was trained exclusively on human speech rather than written text. Block characterises the model as able to support realistic, real‑time conversations by processing audio directly and responding without first converting speech into text.
“Dessa is the world’s first AI you can have a realistic live conversation with that was trained only on human speech,” Block said in its announcement
What makes Dessa different
Traditional voice assistants typically perform a sequence of steps: convert spoken words into text, run language processing over that text, then generate speech. Block says Dessa uses a speech‑to‑speech architecture that lets it listen and speak simultaneously, enabling a continuous conversational exchange rather than discrete request‑and‑response turns. Block also noted that the model was trained on a single NVIDIA GB200 NVL72 system and that it learned language with far fewer words than is typical for text‑based large language models.
Capabilities and limitations
Block says Dessa can attend to conversational cues such as tone, pauses, laughter and sarcasm—signals that text‑centred systems can miss. The company also claims the model can respond in real time without an intermediate text representation. However, Block acknowledges limitations: demonstrations have shown that Dessa can make errors, including misidentifying speakers. Those shortcomings will be material for applications that demand accurate speaker recognition, auditability or safety guarantees.
- Training approach: trained only on human speech, not text
- Architecture: speech‑to‑speech, designed for simultaneous listening and speaking
- Hardware: trained on a single NVIDIA GB200 NVL72 system, according to Block
Why this matters for India
Voice AI advances are significant for the Indian market, where multilingual speech interfaces and low‑literacy accessibility shape product adoption. A model that better interprets prosody and conversational cues could improve automated customer support, voice banking, accessibility features and contact‑centre automation. At the same time, the reported errors underline the need for careful evaluation before deployment in areas involving financial transactions, sensitive personal data or legal decisions.
Block said it plans early access for researchers; the company’s announcement frames Dessa as an alternative to text‑centred training rather than a finished commercial product. The model’s practical value will depend on independent evaluations of accuracy, robustness across languages and dialects, and safeguards for privacy and misuse—areas regulators and enterprises in India will be watching closely.
