Hume's empathic voice interface is one of the few voice products where the difference is obvious within a minute of using it. Turn taking feels natural, latency is low enough that interruption works, and the speech carries prosodic variation that most text to speech does not. The developer documentation gets you from nothing to a working conversation quickly, the API surface is well specified, and the integration guides are the kind of clean technical writing that saves you an afternoon. On the engineering, this is a good resource.
The expression measurement side is where I part company with the material. Hume publishes models that infer emotional states from voice, face and language, and the documentation presents their outputs in confident terms, with named emotions and scores that look like measurements. The underlying science is far less settled than that framing implies. The relationship between observable expression and internal emotional state is contested, varies across individuals and cultures, and is context dependent in ways a per utterance score cannot capture.
None of that uncertainty appears in the documentation, and a developer reading it would reasonably conclude they are receiving a reading of how the user feels. They are receiving a statistical association with training data. That distinction matters enormously once you build on it. A product that routes support tickets, adjusts a sales script or flags a caller as distressed based on these outputs is making consequential decisions on a weak signal, and the failure mode is invisible because the number always looks plausible.
The documentation offers no guidance on validating the outputs against your own population, no discussion of error rates, and no advice on where the signal is too weak to act on. That is the single largest gap. Cross cultural validity goes unexamined and it should not. Vocal expression of emotion varies substantially across languages and cultures, models trained predominantly on one population perform worse on others, and any product deployed internationally will encounter this.
The material treats the models as universal. Consent is the ethical omission. Inferring emotional state from a person's voice is a meaningfully different act from transcribing what they said, and users generally do not expect it. There is no material on disclosure, on regulatory exposure in jurisdictions that treat biometric and emotional inference as sensitive, or on when you simply should not do this.
A vendor is unlikely to write that page, and its absence is still a mark against the resource. Cost behaviour is the practical footnote. Continuous conversational audio consumes differently from discrete requests, and testing with short interactions gives you a misleading sense of what a real deployment costs. Model it on realistic session lengths before committing.
Three point five. Strong technical documentation for a voice product that genuinely advances the state of things, marked down hard for presenting contested emotion science as settled measurement and for leaving every ethical and validity question to the reader.