Text was the first interface for modern AI because text was easy. It was easy to tokenize, easy to score, easy to benchmark. But it was never the natural one. Humans have been speaking for roughly a hundred thousand years and writing for only five thousand. It’s clear that when people are given the choice, they speak.
That choice is now arriving at scale. Customer support, healthcare intake, education, in-car assistants, and consumer companions are all shifting from typed exchanges to spoken ones. Latency has dropped to conversational speeds. Word error rates have fallen to the point where transcription is rarely the bottleneck. The infrastructure works.
And yet anyone who spends real time with voice AI knows the experience is not solved. The systems transcribe accurately and respond quickly, but conversations still feel brittle. Something is missing, and it is not something that shows up in the current metrics most of the AI industry uses.
What voice carries that text does not
A useful way to think about the difference: text is a lossy compression of speech. When a conversation is reduced to a transcript, an enormous amount of signal is discarded. Pitch contours, pacing, pauses, emphasis, breathiness, volume dynamics, and the micro timing of turn-taking all disappear.
That discarded signal is not decorative. Decades of research in psychology and behavioral science show that listeners rely heavily on vocal cues, independent of the words themselves, to judge a speaker's confidence, sincerity, emotional state, and intent. A flat "I'm fine" and a strained "I'm fine" are the same string of characters and two entirely different messages. Humans resolve that difference instantly and unconsciously. It is one of the oldest capabilities we have.
This creates a structural problem for voice AI evaluation. Metrics inherited from the text era, and from the transcription era before it, measure whether a system got the words right. They say almost nothing about whether a system got the conversation right.
The measurement gap is now the capability gap
For most of the history of speech technology, word error rate was a reasonable proxy for quality because getting the words right was genuinely the hard part. That era is ending. On clean audio, leading models transcribe about as well as people do. Several established benchmarks are approaching saturation, which may sound like success but actually signals something else: the tests are no longer measuring the frontier.
The frontier has moved to questions that are harder to formalize. Does the system hear hesitation and slow down? Does it recognize frustration before the user says the word "frustrated"? Does it maintain a consistent vocal identity across a twenty-minute conversation? Does it hold up when the caller has an accent, a bad connection, or a crying child in the background?
These are the qualities that determine whether people trust a voice system enough to keep using it. They are also the qualities most current evaluation practices were never designed to capture. The result is a widening gap between benchmark performance and real-world performance. In this gap, deployments quietly fail.
Why this matters commercially, not just scientifically
The teams shipping voice AI today are discovering an uncomfortable pattern. A model clears every internal quality bar, launches, and then user satisfaction lags projections. Escalation rates stay high. Session lengths stay short. When teams dig into recordings, the failures are rarely transcription errors. They are conversational and emotional failures, including: missed cues, tonal mismatches, responses that were technically correct and socially wrong.
In other words, the industry can now build systems that speak well before it can measure whether they listen well. In every mature engineering discipline, measurement precedes reliable improvement. You cannot systematically fix what you cannot systematically see.
This is the problem we work on at Hume. Our research over the past decade has focused on mapping the structure of emotional expression in the voice, and on building the measurement infrastructure that lets voice AI systems be evaluated on the dimensions humans actually care about: understanding, expression, appropriateness, and reliability under real-world conditions.
The next benchmark era
The history of AI progress is, in large part, a history of evaluation. ImageNet did not just measure computer vision, it organized a decade of it. Shared benchmarks concentrated effort, made progress legible, and turned vague aspirations into tractable engineering problems.
Voice AI is at the point where it needs the same thing, but for the human qualities of conversation rather than the mechanical ones. Building that requires more than a test set. It requires a scientifically grounded account of what emotional and conversational quality actually is, large-scale human judgment to anchor it, and infrastructure that can apply it consistently across models, versions, and production systems.
That is the work ahead for the field, and it is the work we have chosen to focus on. We believe voice is going to be the way most people interact with AI. Whether those interactions feel human will depend on whether we learn to measure what human actually means.
If you are building or deploying voice AI and want to understand how your system performs on the dimensions that matter in production, get in touch. It’s what we do.
