Hume AI

Every voice AI team hits the same wall. The word error rate is good, latency is inside target, the transcripts look fine — and people still hang up.

That gap exists because the metrics measure whether the system was accurate, while users are deciding whether they want to keep talking to it. Those are different questions with different answers.

Hume AI's proposition is to make the second question measurable: put real people behind an API, get scored ratings back in hours, and treat human judgement as infrastructure rather than as a research project.

Hume AI: Measuring Voice The Way People Hear It

• Human raters behind an API — pre-screened participants returning per-sample scores and free-response feedback in hours.
• Kairos builds evaluation suites from real use cases, simulates agent-to-agent and human-to-agent conversations, and tracks regressions.
• Expression measurement across 48-plus emotion categories, 50-plus languages and 600-plus output metrics.
• Public leaderboards across recognition, understanding, text-to-speech and speech-to-speech, ranked by human judgement.
• A leaderboard of the judges — which small models track human ratings most closely, so you know which evaluator to trust.

What Does Emotional Intelligence Mean In Voice AI?

Hume frames voice quality as three layers: reliability, safety and trustworthiness at the base; expressivity, naturalness and flow above it; and emotional intelligence at the peak, described as the dimension that determines whether people actually want to keep talking.

That hierarchy is the most useful thing on the page, because it explains why teams optimise the wrong thing.

The base layer is where all the standard metrics live — recognition accuracy, uptime, refusal behaviour. It is table stakes, it is measurable, and it is what every dashboard shows.

The middle layer is where most complaints actually originate. An agent that is accurate but stilted, or interrupts, or delivers a joke in the same tone as a refund policy, fails on expressivity while scoring perfectly on accuracy.

The top layer is the one nobody can measure with a metric, and Hume states this directly: metrics measure accuracy, while expression, emotion and alignment require human judgement. That is an honest admission from a company selling measurement, and it defines exactly what it is selling.

Metrics tell you the system was accurate. People are deciding whether they want to keep talking. Those are different questions.

Judging The Judges

Alongside four leaderboards ranking voice models, Hume publishes a fifth that is different in kind: not a model leaderboard but a leaderboard of the judges, ranking which small language models track human ratings most closely.

This is the most original idea on the site and worth explaining, because it addresses the central practical problem with human evaluation.

Human ratings are the gold standard and they are slow and expensive. So every team eventually uses a model as an automated judge to score outputs at volume — and then has no idea whether that judge agrees with the humans it replaced.

Measuring the agreement directly, and publishing it, tells you which cheap evaluator is safe to run at scale and which one is quietly rewarding the wrong things.

It is the same insight that makes evaluation frameworks like Promptfoo useful — a test suite is only as good as the assertion behind it — applied one level up, to the assertion itself.

Kairos, Simulation And Regression Tracking

Kairos is described as letting voice teams build evaluation suites from real-world use cases, simulate agent-to-agent and human-to-agent conversations, and track regressions over time.

Simulation is what makes voice evaluation tractable at all, and the reason is arithmetic.

A voice agent's behaviour depends on what the caller says, and callers say an unbounded number of things. Testing by hand means someone calling the system repeatedly, which caps you at dozens of conversations per release and misses the interaction that actually breaks it.

Agent-to-agent simulation removes that cap. One model plays a difficult caller — interrupting, changing its mind, going off-topic — while the system under test responds, and you can run thousands of those.

Regression tracking is the half that matters over time. Voice systems get worse in ways nobody notices: a prompt change fixes one behaviour and breaks turn-taking; a model upgrade improves accuracy and flattens tone. Without a scored suite running every release, you find out from a customer. Designing the agent under test is a separate discipline, handled on a canvas by Voiceflow.

• Separate the three quality layers before deciding what to measure.
• Simulate difficult callers, not cooperative ones.
• Run the evaluation suite every release, not before launch only.
• Check which automated judge tracks human ratings on your task.
• Ask any rating provider how participants are screened.

Leaderboards And Judging The Judges

The Real World VoiceEQ benchmark spans four leaderboards — speech recognition, speech understanding, text-to-speech and speech-to-speech — ranked by human judgement rather than automated scoring.

Publishing a benchmark you also sell services around invites an obvious question, and it is worth addressing rather than ignoring.

The mitigation is that the benchmark ranks other companies' models rather than a product Hume sells into the same comparison, and that the evaluation questions are published. The site lists them plainly: does the voice fit the persona, does the model sustain character voices when reading a story, are jokes and reassurances delivered in distinct tones, can it reliably read out web search results.

Those are unusually concrete evaluation criteria, and publishing them is what makes a benchmark checkable rather than merely authoritative.

The remaining caution applies to any vendor-published benchmark: read the methodology, note who funded it, and treat the ranking as one input. That the criteria are this specific makes it easier to do, not harder.

Human Raters Behind An API

The human feedback product is described as pre-screened participants returning per-sample scores, free-response feedback and aggregated analysis in hours rather than days, through a single API call from study creation to results.

Turning human evaluation into an API call is a bigger operational change than it sounds.

Running a human study conventionally means recruiting, screening, briefing, distributing, chasing, and cleaning the results — days of coordination that no engineering team wants to own, which is why human evaluation gets skipped and automated proxies get used instead.

Collapsing that to one request changes when it happens. Evaluation that takes hours can run inside a development cycle; evaluation that takes days runs before a launch and never again.

Screening and quality monitoring are the parts to interrogate. Paid rating populations attract bad-faith responses, and the difference between a carefully screened panel and a poorly monitored one is the difference between data and noise. Hume names screening and quality monitoring as built in — the questions for procurement are how participants are recruited, what the exclusion criteria are, and how disagreement between raters is handled. None of it replaces the telephony layer, where an open framework like Vocode does the unglamorous work.

Evaluating A Voice Agent Before Shipping

Voice agents fail in production on the dimensions nobody instrumented, and by then the users who left are not coming back to tell you why.

This sequence builds evaluation in the order the layers actually matter.

1. Fix reliability and safety first — nothing above it matters if the base fails.
2. Build an evaluation suite from real transcripts, not imagined scenarios.
3. Simulate difficult callers who interrupt, ramble and change their minds.
4. Get human ratings on a sample before trusting any automated judge.
5. Check which judge model tracks those humans on your specific task.
6. Run the suite every release and watch for regressions in tone, not just accuracy.

Hume AI: Products And Research

Does Your Voice Team Need Human Ratings?

Measure with humans when

• Your metrics look fine and users still abandon conversations.
• Tone, persona or emotional fit is part of the product.
• You need to know whether a release regressed on anything but accuracy.
• You use an automated judge and cannot tell whether to trust it.
• You are shipping voice into a regulated or high-stakes setting.

Automated metrics suffice when

• You are consuming a voice API rather than building the model.
• Word error rate genuinely is the only thing that matters to you.
• The product is text, where evaluation is comparatively settled.
• Nobody would act on the ratings once they existed.
• You are pre-product and have no real transcripts yet.