Our Science
The science behind emotionally intelligent voice AI
Hume's work is based on over a decade of research into human emotion and expression together with our experience building voice models. We turn that foundation into the data, measurements, and feedback researchers need to develop voice AI that listens, speaks, and responds more appropriately.
From emotion theory to computational science
David Hume’s theories of emotion and Charles Darwin’s observations of expression laid early groundwork. Experimental and cross-cultural studies in the 1960s strengthened the field’s scientific foundations.
Today, Hume AI combines computational methods with large-scale human studies to turn that science into data, measurements, and feedback for voice AI development.
Learn moreA history of emotion science
1739
David Hume
Theorizes that emotions drive choice and well-being.
1872
Charles Darwin
Observations of facial, bodily, and vocal expression.
1960s
Emotion research
Researchers undertake experimental and cross-cultural studies on emotional expression.
Today
Hume AI
Hume’s scientists use data-driven methods, computational techniques, and large-scale human studies to understand how people express and perceive emotion across voice, face, and language.
Voice carries more than words
Meaning comes from the sounds we make, how we deliver words, and how listeners interpret them. Research by Hume and our collaborators helps explain these layers and provides a scientific foundation for building voice AI.
Vocal bursts
Brief, nonverbal sounds such as laughs, sighs, cries, and gasps convey a range of emotional meanings. U.S. English-speaking listeners distinguished at least 24 meanings in these sounds.
Read the study(opens in a new tab)Prosody
Prosody is the pitch, pace, rhythm, loudness, and emphasis in speech. With words held constant, English-speaking listeners in the U.S. and India distinguished at least 12 shared emotional meanings.
Read the study(opens in a new tab)Cross-cultural perception
People across cultures interpret vocal expression in both shared and distinct ways. A study across five countries found that vocal bursts conveyed both shared and culture-specific meanings.
Read the study(opens in a new tab)
Mapping the meanings of expression
Semantic Space Theory combines large-scale human studies with statistical modeling to map the emotional meanings people perceive in expression, and how those meanings blend.
Hume uses this framework to define expression categories, annotate data, and evaluate voice models.
Learn more about Semantic Space theoryLoading Hume’s semantic-space map…
We made the science usable for model development
Building expressive voice models required us to translate studies of human perception into engineering decisions: what to represent, which data to collect, how to annotate it, and how to tell whether a model improved.
Represent expression
Our detailed taxonomy represents combinations such as quiet anger or restrained excitement that simpler labels can miss.
Explore our 600+ tags600+Tags
- Emotion categories
- Afraid, Angry, Amused, Joyful, Disgusted, Excited, Distressed, Sad, Surprised
- Speaking styles
- Whisper, Calm, Excited, Nervous, Assertive
- Vocal qualities
- Pace, Warmth, Energy, Clarity, Hesitation
Build the right data
Our cross-product sampling research explores expressive combinations that ordinary filtering can miss, giving models richer examples to learn from.
Explore cross-product sampling- Quiet + angerExpression × delivery
- Quiet + excitementExpression × delivery
- Forceful + angerExpression × delivery
- Forceful + excitementExpression × delivery
Conceptual illustration of sampling combinations
Design useful measurements
Our Real-World VoiceEQ benchmark examines voice performance through human-grounded evaluations and compares automated judgments with human ratings.
Read the benchmark whitepaper(opens in a new tab)- Speech generation
- Conversation
- Understanding
- Transcription
- 01Defined tasks
- 02Listening rubrics
- 03Scoring methods
Connect feedback to model development
For a major frontier lab, we developed bespoke evaluations of multi-speaker dialogue to identify priorities for the next model iteration.
Read the frontier-lab case study
Bring your next voice research question to Hume
Work with researchers and engineers who connect the science of human expression to your model-development goals. We help define the right evaluations, identify improvement priorities, and measure progress across model generations.