Speech AI model Dessa from Block reads emotion directly from audio

What happened

Block, the company behind Square and Cash App, has introduced Dessa, a speech model that processes audio directly rather than converting it to text first. According to The Decoder, Dessa is designed to recognise emotion, sarcasm and pauses in real time.

The model reportedly runs locally on a MacBook Pro M4 Max, and it is not yet publicly available. The headline framing is that Dessa was trained entirely on human speech, with the aim of catching signals that text-based models tend to miss.

At the time of writing, the story rests on a single report from The Decoder. Technical detail on training methods, evaluation results and a release timeline has not been published in the material available.

Why it matters

Most voice systems in production still follow a transcribe-then-interpret path. Converting speech to text removes tone, hesitation, emphasis and timing, which are often the cues that tell a listener whether a customer is frustrated, confused or being sarcastic. A model that reads the audio itself could, in principle, give downstream systems richer input about how something was said, not only what was said.

For leaders in experience and digital transformation, the significance lies in the architecture. If a speech model can run on local hardware, organisations may have more options for handling sensitive audio without sending it to a remote service. That changes how teams might weigh privacy, latency and cost when designing voice-based service journeys.

The Renascence take

The interesting question is not whether Dessa can detect emotion, but whether an operator can act on that signal responsibly. Emotion detection is only useful when it is tied to a clear service decision, such as when to escalate, when to slow down or when to offer a human. Without that link, it becomes a dashboard metric that nobody trusts.

Emotional cues in voice are valuable only if they change what happens next for the customer. Most organisations will focus on the detection accuracy and miss the design question of which intervention a detected frustration should trigger. A customer-obsessed operator should map those interventions before buying the model, and treat any emotional inference as a prompt for a human-centred response, not a verdict on the caller.