NEW YORK — Block, the financial technology company led by Jack Dorsey, recently unveiled Dessa, a new artificial intelligence (AI) voice model, according to reports from outlets like The Economic Times and IndiaTimes. The company describes Dessa as a full-duplex speech-to-speech AI model, uniquely trained from scratch solely on human conversation, using audio data exclusively and bypassing text or pre-trained models.
This novel training approach distinguishes Dessa from conventional voice AI systems. Instead of converting speech to text and then processing it, Dessa directly processes audio, a method Block states preserves emotional cues and variations in tone that are often lost during text conversion. This allows the AI to recognize and express emotions such as sarcasm, laughter, frustration, and sadness through natural voice interactions, as reported by Newscord.org.
Block states that Dessa can independently determine what to say, when to respond, and how to express itself, including its voice and conversational style. Demonstrations have shown the model reacting naturally to humor by laughing mid-sentence and adjusting its speaking volume when asked, behaviors it reportedly developed spontaneously without explicit training for them. The-Decoder.com highlighted that Dessa's inference runs faster than real-time, enabling replies without the awkward pauses often associated with current AI voice assistants.
The development of Dessa was motivated by a blend of curiosity, a desire for improved human-AI communication, and safety concerns regarding the inability of existing voice AI to register emotion, according to Block's statements. The company claims Dessa learned language from "hundreds of times fewer words" than large language models (LLMs), a significant efficiency gain if proven broadly applicable.
Block claims Dessa stands as "the most authentic full-duplex speech model ever built," asserting that "talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence." The company also states that, "as far as it knows, Dessa is the world's first AI you can have a realistic live conversation with that was trained only from human speech." These claims, while compelling, are currently presented by Block as allegations, awaiting broader independent verification.
Ragavan Thurairatnam, the lab's lead, described Dessa as "a GPT3 moment for speech," suggesting a foundational shift in AI capabilities. Early demonstrations have been described by some as sounding "genuinely empathetic" in life advice recordings. User comments on platforms like X, such as Alexander Hoff's remark that he had never "heard something artificial that sounded so… human," and Hesamation's description of Dessa as "pure science fiction," reflect the initial impression of its advanced realism.
Technically, Dessa was trained on a single NVIDIA GB200 NVL72 system, and Block has indicated that the model can run locally on a MacBook Pro M4 Max or better, as noted by Notebookcheck.net. This local operability could have implications for privacy and accessibility, allowing advanced AI interactions without constant cloud connectivity.
Recognizing the potential for new technologies, Block is providing early access to Dessa to a small group of AI safety researchers to examine potential risks associated with the new model. The specific "potential risks" that these researchers might identify, however, remain unknown. This proactive approach underscores the company's stated commitment to safety, even as the full implications of such a realistic voice AI are still being explored.
Many aspects of Dessa's capabilities and impact are still unknown or only forecasts. Researchers are keen to understand how efficiently models trained solely on speech, like Dessa, will learn language and facts compared to text-trained models. The efficiency with which Dessa can learn new languages and the sorts of in-context learning it can perform are also open questions. Block has not publicly disclosed the model's size or benchmark results, nor has a public release date for Dessa been announced.
The full utility of Dessa, particularly concerning its broad world knowledge and tool calling integration, remains a subject of discussion within the AI community. Beyond technical performance, deeper philosophical and ethical questions persist. Whether recognizing a tone of urgency, fear, or pain in a voice pattern is equivalent to the AI genuinely sympathizing with the underlying emotions is a complex issue. The long-term implications of highly convincing AI voices, including their potential for fraud, imitation, or even fostering unhealthy comfort, are significant concerns.
Filmmaker Justine Bateman has publicly questioned why AI voices should sound human at all, arguing they could be used to deceive people, a sentiment that resonates with broader societal anxieties about AI authenticity. While Block aims to "bridge the gap between human communication and AI understanding" with Dessa, the extent to which this gap will be truly closed, and the consequences of doing so, are subjects that will unfold over time. A public waitlist for Dessa is currently open, signaling Block's intent to eventually broaden access to this innovative AI voice model.