Disa model breaks text mediation with a full voice conversation built from scratch using dedicated training hardware

Audio interaction with AI has remained constrained by an outdated technical compromise: extracting human voice intonations into abstract textual words, feeding the model with them, and then reconstructing them into speech, a trick that drops emotion, hesitation, and sarcastic intonation cues. The AI lab at Block broke this loop by introducing the conversation model Disa, a fully dual inference system trained from scratch on audio recordings of human conversations only, without relying on written texts or pre-trained models.

The new model treats audio as a direct input and immediate outputallowing exchange of speech and listening at the same moment, like natural conversation between people. Disa does not wait for a sentence to finish before responding; it can interrupt spontaneously and produce immediate interaction cues such as "yes" and "got it", as well as laugh in the middle of a phrase when it senses a joke. Notably, our observation of the infrastructure shows that the team, led by Ragavan Thooriratnam and reporting directly to Jack Dorsy, trained the model entirely on a single Nvidia GPU-200 NVL 72 compute rack, enabling local inference to run faster than real time on personal computers equipped with M4 Max processors.

The technical paradox lies in data economics; massive language models consume billions of textual words and must learn programming code alongside natural languages, while Disa learns conversation with a vocabulary that is hundreds of times smaller. This is due to the enormous density of raw audio, where a single spoken word occupies thousands of times more bytes than its written counterpart, including breaths, intonation and speech rate. The experiment shows that a sentence’s meaning can change entirely depending on the word that is emphasized vocally, an expressive dimension that textual transcription erases completely.

Moving beyond textual mediation touches the core of safety standards for intelligent systemsCurrent systems that rely on speech-to-text become functionally blind to cries of pain or sudden emotional changes in the speaker, and continue to reply with textual coldness, ignoring the urgency of the situation. This challenge led Block’s lab to limit early access to the model to AI safety researchers to uncover unexpected risks of a system that learns directly from the ear, especially after experiments that compared its performance with external dialogue models such as GPT-Live 1 from a third-party provider.

When examining this architectural shift, it recalibrates the engineering and investment calculations for businesses in the Gulf, Egypt and the Levant, particularly in contact centers and automated banking and government services. Regional institutions have long relied on a three-stage technology chain that starts with voice recognition, then calls cloud-based language models, and finally generates speech, which inflates response times and cloud-computing bills, as well as weakening local dialects and stripping them of their communicative warmth. Running direct speech models locally on advanced desktop computers allows banks and critical agencies to reduce monthly subscription costs while ensuring data sovereignty by processing customer voices on on-premise servers instead of transmitting them across borders.

On the development side, this leap imposes a new reality on AI teams in Riyadh, Dubai and Cairo; building voice assistants that can understand Arabic dialects is no longer tied to gathering digital texts or digitizing articles, but to collecting and indexing real audio recordings that reflect verbal overlap, vocal intonation and spontaneous dialogue interruptions in everyday Arab environments. The next race for those designing automated interaction tools in the region is gradually shifting from engineering textual commands to engineering raw audio and its field-level emotional contexts.