Wire flash
Hume AI releases audio-native API to detect caller emotion beyond transcripts
Editorial responsibility
- No named human review is recorded for this page.
- Source reporting is collected, normalized, translated or condensed automatically when needed.
- Automatically published source-backed update
The post argues that voice agents and QA dashboards grade calls on transcripts, which discard tone, hesitation, and interruption cues, so angry callers can score green. It says speech-to-text and synthetic voices have advanced, but the agent in between still runs on text, and teams optimize what the transcript captures. Hume AI (@hume_ai) has released an API that answers plain-English questions about audio, such as whether a caller sounds unsure, returning a probability per answer. The API is audio-native, so it listens to the recording itself without a separate classifier per use case. Hume says it is trained on millions of human judgments and, in initial evals, matched a panel of human listeners more closely than a single human rater did, which the post says could let a lab score every model checkpoint instead of waiting on a panel.
Source report
You've yelled "representative" at a phone three times, a little louder each time, and heard "I can help with that!" three times back. On that company's QA dashboard, the call scored green. The bot was reading a transcript, and a transcript can't hear you losing it.
The Same Blind Spot, Everywhere
The problem isn't isolated. Consider what gets lost when a voice agent only reads text:
- You say "I'm fine" in a tone that means the opposite, and the bot takes you at your word.
- A hesitant "...yes" to a payment confirmation gets booked as a confident one.
- If you pause to think, the bot decides you're done and talks right over you.
Speech-to-text can transcribe almost anything now, and synthetic voices sound strikingly human—yet the agent sitting in between still runs on text. A furious caller and a calm one can look identical on the page.
Why Nobody Fixed It
The tooling can only read the transcript, so teams optimize the transcript—and the angry customers who hang up live in the part nobody kept.
A New Audio-Native Approach
@hume_ai has just released an API that goes after exactly this problem:
- You ask it a plain-English question about the audio ("does this caller sound unsure?") and get back a probability per answer.
- It's audio-native, so it listens to the recording itself—no separate classifier needs to be built for each use case.
- Hume says it's trained on millions of human judgments. In their initial evals, it matched a panel of human listeners more closely than a single human rater did—which would let a lab score every model checkpoint instead of waiting on a panel.
The Bottom Line
Voice agents have spent years getting graded on the transcript—the one file that throws away how the caller actually sounded.
Source
aakashguptaNeutral / independent
Part of this Story
Hume AI launches audio-native Decision API for emotion and speaker analysis