Hume AI

Hume AI grew out of a research program in psychology. With Dacher Keltner (UC Berkeley), Alan Cowen developed “semantic space theory”: the idea that the experience and expression of emotion are organized into many categories (27 for reported experience, 24 for vocal emotion, 28 for facial expressions), linked by continuous gradients rather than reduced to a handful of basic emotions (Cowen & Keltner, 2017, 2021).

This theory sits between two classic positions: basic emotion theory (Ekman) and constructionism (Barrett). The company derived its commercial models from it, which report several dozen “dimensions” of expression. It claims a lineage from David Hume, after whom it is named, Darwin and Ekman.

EVI (Empathic Voice Interface) is a voice conversational agent. You speak to it and it answers aloud, in real time. It analyzes the way you speak (tone, rhythm, timbre) and adjusts the tone of its reply: more apologetic in the face of frustration, more sympathetic in the face of sadness. It detects the end of a speaking turn and can be interrupted. The content of its replies is produced by a third-party language model (Anthropic, OpenAI, Google, among others), chosen by the developer.

Octave is a text-to-speech system that adapts pronunciation, pace and emphasis to the meaning of the text. It can design a voice from a description, and clone a voice from 15 seconds of recording.

On January 22, 2026, Alan Cowen and about seven engineers joined Google DeepMind under a non-exclusive licensing agreement. This is not an acquisition: Hume AI continues to operate independently, led by Andrew Ettinger, who comes from the AI data sector.

The company now presents itself as a “data and evaluation layer” for voice AI: it sells evaluations and human judgments to labs developing voice models, and publishes a ranking of the “emotional intelligence” of voices. In June 2026 it discontinued its flagship expression-measurement API, the product that had built its reputation. Its home page nevertheless still features it at the time of writing.

As early as 2023, Hume AI announced research collaborations with US hospital systems and targeted screening, mood monitoring and triage. In 2024 it presented its conversational voice as a possible support against loneliness and a complement to therapy in the face of a shortage of clinicians.

Hume AI deserves a card on this site for two reasons. The first is scientific: rarely has an AI company come so directly out of a research program in the psychology of emotion. Its story stages a debate psychologists know well, about the link between expression and emotion, and reminds us that a voice that sounds empathic is not a voice that understands.

The second is practical. Patients increasingly encounter these voices without naming them: they are built into coaching, well-being or companionship apps that do not always disclose their provider. The REAM example shows what a clinician gains by checking: who processes the voice, where the data go, what framework of use is claimed, and whether the sales pitch matches the terms of use.

Finally, the 2026 reorganization is a reminder of how fragile this sector is: a tool cited in a study can disappear or change its nature within months. Recommending a voice app to a patient requires knowing its publisher, not just its voice.