A radical approach: native speech intelligence
On October 23, 2026, Block announced the launch of Dessa, an artificial intelligence model designed specifically for voice interaction. Unlike most current assistants, which rely on a processing chain that converts audio to text, processes it, and then synthesizes it back, Dessa operates as an end-to-end voice-to-voice system. This means the model receives audio and produces audio without passing through an intermediate textual transcription stage.
The company describes this approach as an experiment in 'native speech intelligence'. The premise is that, just as humans learn languages by listening before reading, a model can acquire linguistic and conversational competencies directly from acoustic signals. Block deliberately excluded any pre-training based on books, web pages, or text datasets to explore what capabilities emerge when a model is exposed only to the dynamics of human conversation.
Full-duplex architecture and dialogue management
One of the most notable technical features of Dessa is its ability to operate in full-duplex mode. In practice, this allows the system to listen and speak simultaneously, mimicking the behavior of a real telephone call. The model can react to user interruptions, interject acknowledgment sounds (such as 'yes' or 'uh-huh') while the interlocutor is speaking, and adjust its pace to match the conversation.
This architecture allows the model to capture non-verbal signals that are crucial for natural communication. According to Block's materials, Dessa manages the rhythm and emotional aspects of dialogue, including the detection of sarcasm, irony, or changes in the speaker's emotional state. Public demonstrations showed the model laughing mid-sentence, raising its volume when requested, and completing phrases, although occasional errors were also observed, such as incorrectly addressing a user by a foreign name.
Comparison of voice processing flows
Table contrasting the traditional approach with that of Dessa.
| Feature | Traditional Model (STT-TTS) | Dessa (Speech-to-Speech) |
|---|---|---|
| Intermediate processing | Transcription to text | None (direct audio) |
| Interruption management | Limited (waits for user to finish) | Full-duplex (real-time reaction) |
| Nuance detection | Based on keywords | Based on tone, pauses, and rhythm |
| Perceived latency | Higher (multiple stages) | Lower (continuous process) |
Exclusive training with human audio
Block's research team, led by Ragavan Thurairatnam, trained Dessa using exclusively human conversation audio. The company states that no text data or pre-trained models in written language were used. This method seeks to replicate human language acquisition, where auditory input is the primary source of learning before exposure to writing.
According to the company's statements, this approach allows the model to learn language structure, social conversation norms, and emotional signals directly from the source. However, analysts point out that this methodology raises open questions about the efficiency with which the model acquires specific facts or structured knowledge, areas where text-based models usually have a clear advantage due to the information density in written corpora.
Hardware requirements and local execution
Block has highlighted the computational efficiency of Dessa. The company claims that the model's training was performed on a single NVIDIA GB200 NVL72 rack, a configuration significantly smaller than that used to train frontier text-based language models, which typically require tens of thousands of accelerators.
Regarding inference, Block indicates that Dessa runs faster than real-time on a MacBook Pro M4 Max or higher. This suggests that the model has a parameter size and memory footprint that allow its execution on high-end consumer hardware. The ability to run the model locally has important implications for privacy, as conversations do not need to be sent to external servers for processing.
Advantages of local inference
Local execution process of Dessa.
- 01The user speaks and the audio is captured locally.
- 02The Dessa model processes the audio on the device's CPU/GPU.
- 03The audio response is generated without sending data to the cloud.
- 04The response is played in real time, maintaining low latency.
Restricted access and risk assessment
Despite the announcement, Dessa is not available to the general public. Block has limited early access to a small group of AI security researchers. The company has stated that its goal is to study the implications and risks associated with a voice model that sounds convincingly human and reacts to emotional signals.
The main concerns revolve around identity impersonation and trust manipulation. Unlike text chatbots, a voice that can detect and respond to emotional nuances has a greater potential to be used in frauds or to influence users' decision-making. Block has indicated that it wants an external evaluation of the model's behavior before a wider distribution, acknowledging that public demonstrations, while promising, do not constitute a sufficient basis to assert the general reliability of the system.
Open questions
- The claim that training was performed on a single NVIDIA GB200 NVL72 rack has not been independently verified.
- The model's ability to retain factual and structured information without text pre-training is not confirmed by external evaluations.
- The exact response latency in milliseconds has not been published, so the claim of 'faster than real-time' is not reproducible with precise data.
- There is no public API or detailed technical documentation that allows third parties to evaluate the model's internal architecture.
Keep exploring
Sources consulted
Corrections and transparency
If you spot incorrect or outdated information, send us a correction with the page and source we should review.
Submit a correction