Hume AI lets you query raw audio for tone and emotion

The most telling moment in a customer service call isn't usually what the caller says. It's the sharp intake of breath, the awkward silence, or the resigned sigh right before they hang up.

Yet for years, nearly every voice AI and automated quality-assurance tool threw that entire layer of information away. Audio got flattened into text, and in plain text, "I'm fine, thanks" looks identical whether the customer is genuinely satisfied or furious.

Hume AI unveiled its Audio Decision API. Instead of converting speech to text first, it lets systems run custom queries directly against raw sound.

Ask your audio a direct question

Traditionally, changing what a call center scores is a messy ordeal. You write new rubrics, build new pipelines, wait weeks for evaluations, or spend real money fine-tuning specialized classifier models.

Hume replaces that setup with a single query. You stream in raw audio, define a question along with possible answers in a line of JSON, and get back structured probabilities in real time. There is no retraining and no manual labeling required.

The API handles four distinct formats. You can ask yes/no questions like whether a customer sounds frustrated, multiple-choice questions to categorize agent tone, rating scales to gauge patience, or span queries that pinpoint the exact timestamps of a sigh or a laugh.

It can also track acoustic facts in the same pass. Teams can ask how many speakers are on the line, flag background sirens, or verify whether a voice is AI-generated.

40,000 recordings searched in minutes

Where Hume's approach really pulls ahead is raw throughput.

The company claims the model is fast enough to run dozens of questions every few seconds on a live call, or scan an archive of 40,000 recordings in minutes. That effectively turns a massive, unindexed audio folder into a searchable database without requiring transcription first.

Under the hood, the system leans on Hume's extensive research into vocal perception, backed by tens of millions of human ratings. In evaluations, the Audio Decision model scored a 70.7% agreement rate with average human panel judgments, well above the 56.5% baseline scored by individual human raters.

Hume claims the model outperforms existing audio processing setups across accuracy, speed, and cost. Developers can currently join a waitlist or experiment with both file uploads and streaming audio in Hume's playground.

Why it matters

We think Hume found an elegant compromise between rigid audio classifiers and bulky generative models.

General multimodal models are great at conversational banter, but asking them to generate paragraph-long descriptions just to decide whether an agent should intervene is slow, expensive, and prone to hallucinations. Hume gives you a structured probability instead. That makes it easy to wire directly into business logic: if caller frustration crosses 0.8, hand off the session to a senior rep immediately.

For voice agent builders, this solves a chronic blind spot. Cascaded architectures usually feed text transcripts into an LLM, meaning the bot cannot hear hesitation, sarcasm, or genuine warmth. Giving an agent access to emotional telemetry alongside the transcript makes real-time interactions feel significantly more intuitive.

The catch will be how well these perceptions travel across accents, regional idioms, and languages. Tone is deeply subjective, and what sounds like brisk efficiency in one culture might register as hostile rudeness in another. Teams will need to sanity-check these probabilities before letting them trigger high-stakes operational rules.