Dessa AI Model Masters Conversation Without Text Training
Block, the technology firm behind Square and Cash App, has unveiled Dessa, an innovative AI model designed to engage in human-like conversations by relying solely on audio data. Unlike conventional voice assistants that depend heavily on text processing, Dessa learns from direct exposure to spoken dialogue.
Unique Approach to Speech and Conversation
Dessa operates as a full-duplex speech-to-speech system, enabling it to listen and talk simultaneously. This feature allows it to interject, respond, and acknowledge conversational cues in real-time, mimicking the dynamics of natural human interaction. Instead of incorporating any written text or pretrained language models, Dessa was programmed entirely from scratch using only audio recordings of people engaging in conversations.
This methodology enables the model to grasp subtle speech elements such as sarcasm, emotional tone, laughter, and hesitations, aspects that traditional speech-to-text technologies often miss. Demonstrations of Dessa reveal its ability to laugh mid-sentence, adjust speaking volume when prompted, and seamlessly complete sentences, reflecting a deeper understanding of conversational context.
Technical Insights and Availability
Training the model required substantial computational power, utilizing a single Nvidia GB200 NVL72 system equipped with 72 Blackwell GPUs. Despite its complexity, Dessa is optimized to run efficiently on consumer hardware such as the MacBook Pro featuring an Apple M4 Max chip or superior processors. Moreover, the model processes speech faster than real time, ensuring fluid conversations without perceptible delays.
Currently, Dessa is not available for public use. Block has initiated a limited testing phase by inviting selected AI safety researchers, with a broader rollout contingent on further evaluations. Detailed technical specifics, including parameter size and benchmarking results, have yet to be disclosed by the company.
Implications for Pakistan’s Tech Landscape
The development of Dessa signals a significant advancement in AI-driven speech technologies, particularly relevant as Pakistan seeks to enhance its digital communication infrastructure. As voice-enabled interfaces gain prominence globally, models like Dessa could revolutionize user engagement by providing more natural, emotionally intelligent interactions. This innovation may also inspire local tech firms to explore audio-first AI applications, potentially boosting Pakistan’s role in global AI research and development.
Block’s approach of bypassing text-based training challenges conventional models and highlights new possibilities for creating more empathetic and responsive AI systems. For the Pakistani audience, this advancement underscores the growing impact of AI on everyday technologies, offering a glimpse into the future of digital conversation tools.
“Dessa can interrupt and respond in real time,” said a spokesperson from Block, emphasizing its human-like conversational capabilities.
