What Dimensions Does Hume AI Measure in Voice AI?

Updated

Hume AI measures voice AI across the dimensions people actually experience in a conversation, not just whether a task was completed. According to Hume, "a task can succeed while the conversation fails," so its evaluation platform scores timing, tone, emotional expression, interruption, accent, context, and adaptation. The specific dimensions shown on its site are: Timing, Tone, Emotion, Interruption, Background noise, Accent, Language, Context, Adaptation, and Intent. These are presented as illustrative example data, not real model results.

The measurement gap Hume is addressing

Hume's framing is that "there is more to voice than words." A voice system can resolve the user's request while still failing the interaction — talking over the user, sounding flat when the user is frustrated, or breaking down on an unfamiliar accent. Hume's platform is built to capture those failures, not only task success.

The site poses the gap as two questions:

  • Does the system understand the user?
  • Does the user feel understood?

The dimensions below are how Hume operationalizes the second question.

The dimensions Hume measures

Dimension What it captures Example values shown on the site
Timing Pacing of turn-taking 18% overlap / 240 ms gap; 2% overlap / 1.4 s gap
Tone Vocal quality Warm, Clipped, Measured, Urgent, Flat, Gentle, Bright, Firm
Emotion Emotional expression, scored Amusement 0.62, Frustration 0.92, Anxiety 0.79, Relief 0.46
Interruption Barge-ins during the conversation 7 barge-ins; 0 barge-ins; 11 barge-ins
Background noise Signal-to-noise ratio (SNR) SNR 6 dB; SNR 31 dB; SNR 4 dB
Accent Accent variety handled Midwest US, Glaswegian, Lagos, Chilean, Kansai, Québécois, Punjabi, Estuary
Language Languages covered English, Spanish, Mandarin, Hindi, Arabic, Portuguese, French, Yoruba
Context Conversation situation Follow-up, First call, Escalation, Handover, Renewal, Complaint, Check-in, Repeat
Adaptation Whether the system shifts across turns 2 turns / No shift; 4 turns / 1 turn; 3 turns / No shift
Intent What the user is trying to do Dispute, Confirm, Cancel, Ask, Escalate, Reschedule, Verify, Complain

Timing and interruption

Timing is measured as overlap percentage and gap length in milliseconds — for example, 18% overlap with a 240 ms gap versus 2% overlap with a 1.4 s gap. Interruption is counted as barge-ins, ranging in the examples from 0 to 11. Together these describe whether the system and user are taking turns smoothly or talking past each other.

Tone and emotion

Tone is labeled with descriptive qualities (Warm, Clipped, Urgent, Flat, and so on). Emotion is scored numerically per emotion — Amusement, Annoyance, Anxiety, Calm, Relief, Frustration, Joy, and Concern appear in the example data, with values like Frustration 0.92 and Relief 0.46. This is the layer that separates "the words were right" from "the delivery was right."

Accent, language, and background noise

Accent and Language test coverage across real speaker variation — the site lists accents from Midwest US to Glaswegian to Kansai, and languages from English to Yoruba. Background noise is expressed as SNR in decibels, from a noisy 4 dB to a clean 31 dB. These dimensions test robustness rather than conversational quality alone.

Context, adaptation, and intent

Context describes the situation the conversation sits in (First call, Escalation, Complaint, Renewal, and others). Adaptation tracks whether the system changes behavior across turns — for example, "4 turns / 1 turn" versus "3 turns / No shift." Intent captures what the user is trying to accomplish (Dispute, Confirm, Cancel, Ask, Escalate, Reschedule, Verify, Complain). These three test whether the system reads the situation, not just the sentence.

How these dimensions fit into the evaluation workflow

Hume describes a four-stage loop: Generate → Stress-test → Diagnose → Validate.

  1. Generate — VoiceEQ analyzes your conversation data to identify use cases, generate test cases, and define scenarios and success criteria. Outputs include a criteria list, metric set, and experiment plan.
  2. Stress-test — run the voice system against those cases under realistic conditions.
  3. Diagnose — find why failures happen, using the dimensions above.
  4. Validate — confirm that changes actually improved performance.

The platform is described as covering both models and agents, and combining objective measures with human judgment. Hume states it is trusted by leading AI labs (5 of 7) and cites 50+ languages supported, 100M+ speech samples, and 10+ years of emotion science research.

What to check before relying on these numbers

The model comparison table on the page is explicitly labeled "Illustrative, not real data," so the values above show what the dimensions look like, not how any named model performs. If you want to use Hume's evaluation for your own system, the practical starting point is the Generate stage — feeding in your real conversation data so the test suite reflects your actual use cases, accents, and failure modes rather than a generic benchmark.