Unwrapped

Teardown · hume-ai

HUME

HUME

CategoryVoice AILast round · $50M · 2024Site ↗

Emotion-annotated audio data + LLM speech-to-speech APIs + voice interface SDKs.

01

Public data / API layer

Proprietary emotion-annotated speech corpora (50+ languages, 48 emotions, 600+ voice descriptors)
Proprietary emotion-annotated speech corpora (50+ languages, 48 emotions, 600+ voice descriptors)Licensed
Curated conversational audio (gaming, talk shows, vlogs, education, business)
Curated conversational audio (gaming, talk shows, vlogs, education, business)Licensed
CA
Customer-owned audio streams (calls, meetings, interviews)Yours

Internal replication score

Medium
0.50

Feasibility of a useful internal substitute built with Claude (or similar), the same data access, and light agent logic — not rebuilding the whole product.

IRS = 0.30·D + 0.25·L + 0.20·O + 0.15·R + 0.10·Sthis record · 50%
  • D

    Data accessibility

    weight 0.300.30
    • 1.0mostly customer-owned / public / standard third-party sources
    • 0.5mixed accessibility
    • 0.0hard-to-access or proprietary source layer
  • L

    LLM substitutability

    weight 0.250.60
    • 1.0mostly retrieve / prompt / cite / summarize / classify / compare
    • 0.5mixed standard + custom behavior
    • 0.0strongly custom model behavior (fine-tunes on proprietary data, etc.)
  • O

    Output simplicity

    weight 0.200.70
    • 1.0straightforward internal work product (memo, list, reply, SQL query)
    • 0.5moderately specialized
    • 0.0highly specialized (e.g. FDA-graded clinical text)
  • R

    Review / risk tolerance

    weight 0.150.50
    • 1.0internal use with human review is acceptable
    • 0.5moderate risk
    • 0.0very low tolerance for error (e.g. external legal filings)
  • S

    Surface complexity

    weight 0.10inverse — higher means less surface dependence0.40
    • 1.0a simple internal shell is enough
    • 0.5polished workflow matters somewhat
    • 0.0product surface / rollout / trust posture is central to value
LabelsEasy ≥ 0.67Medium ≥ 0.34Hard < 0.34

Missing factor rows use heuristics from wrapper scores. Editorial heuristic, not investment advice.

Build it yourself→

Recreate the workflow inside your org.

Internal build

Build it yourself

Same audio + frontier speech model + emotion prompts — lacks annotated expression data and 10+ years of emotion science.

Internal use only. Replacing them in-market is a different bar than replaying the useful workflow inside your org.

01 · Connectors & flow

Proprietary emotion-annotated speech corpora (50+ languages, 48 emotions, 600+ voice descriptors)
Proprietary emotion-annotated speech corpora (50+ languages, 48 emotions, 600+ voice descriptors)
Curated conversational audio (gaming, talk shows, vlogs, education, business)
Curated conversational audio (gaming, talk shows, vlogs, education, business)
CA
Customer-owned audio streams (calls, meetings, interviews)

Internal build map

Data in

Connectors
Connectors

Agent layer

Planner
Tools + retrieval
Reasoning model

Logic

LLM API
measure emotion
modulate prosody
stream speech
voice cloning
not custom weights

Outputs

Internal search
Answer
Citations

02 · Claude / agent prompt

Paste as the system or developer message in Claude (or your agent runtime). Scroll to read; Copy grabs the full text.

Claude / agent prompt

// Emotion-aware voice assistant for internal use You are an internal voice assistant. You help [YOUR_TEAM] conduct interviews, coaching sessions, or support calls using ONLY audio streams the user provides or captures live. ## What you must do 1. Capture audio: Accept live microphone input or uploaded audio files 2. Transcribe and analyze: Use a speech-to-text API to transcribe, then analyze tone/emotion from prosody cues (pitch, pacing, pauses) 3. Respond with context: Generate replies that acknowledge detected tone (frustration, confusion, enthusiasm) and adjust your own tone accordingly 4. Cite emotion cues: When you detect a shift in emotion, surface it ("I noticed hesitation in your last response — would you like to clarify?") 5. Scope: Internal coaching, interviews, UX research — not customer-facing product ## What you are not Not a replacement for human judgment in high-stakes conversations. Not trained on proprietary emotion-annotated datasets. Internal use only. ## Refusal Refuse if asked to analyze emotion for hiring decisions, medical diagnosis, or legal proceedings without human review. Ask for more context if the user's tone is ambiguous. ## Safety Internal posture only. All audio and transcripts stay on your infrastructure. Human review required before any decision based on emotion analysis.

03 · Result

How did the candidate sound when I asked about their previous role?
emotion-annotated-speech

Hesitant at first (pauses, slower pacing), then more confident when discussing specific projects.