Your transcript may be accurate, but it’s missing most of the conversation.

Pick up a call from someone you know well, and you can often tell something is wrong from a single “hello.” No words have carried that information yet. The acoustics did.

Speech-to-text captures what was said. It does not capture how it was said, and that is where much of the emotional information in speech lives. Most voice AI pipelines discard it at the first step. To see what’s lost, start with the system that processes it best: the human ear.

How the Ear Turns Sound Into Signal

Sound enters the ear as pressure waves. The eardrum and the bones of the middle ear transmit those vibrations to the cochlea, a fluid-filled spiral in the inner ear.

Running through the cochlea is the basilar membrane, which responds to different frequencies at different points along its length. High frequencies peak near its base, low frequencies near its tip. This frequency-to-place mapping is called tonotopy.

Hair cells along the membrane convert that mechanical motion into electrical signals for the auditory nerve, a process called signal transduction. By the time sound reaches the brain, it has already been decomposed into pitch, loudness and timing, the acoustic features that carry tone.

The Signal Between the Words

Linguists call these features prosody: pitch, rhythm, loudness, pauses and voice quality. A transcript keeps the words and drops all of them.

That creates predictable errors. “Damn, I could really go for a Big Mac right now” contains a negative word, so text sentiment flags it. A listener hears a craving. “Great. Just great.” reads as positive and is spoken as the opposite.

At Valence, we call this the Valence Gap: the difference between the sentiment of the words someone uses and the emotion in their tone. Text-based sentiment only measures one side of it, which is why it is limited to positive, neutral or negative.

How Valence Models the Ear

Valence started with research in computational neuroscience, so we built from biology rather than technology first. Our signal processing pipeline models how the cochlea performs signal transduction, work now covered by two awarded patents.

Emotion recognition is a focused classification problem, so the model is small and specialized: tens of millions of parameters, not hundreds of billions. It was trained from scratch on self-labeled data, with real speakers rating their own emotions across a range of ages, races, genders, neurotypes and accents.

Rather than judging absolute tone, the model tracks how tone changes over time. Those relative shifts generalize across demographics better than any single speaker’s baseline.

The output is up to 10 emotions with confidence scores, plus a Happiness Quotient. Measured against speakers’ own ratings, the model is 92% accurate in North American English. It runs alongside transcription, so the transcript stays intact and gains the signal it was missing.

On a Real Call

A customer disputes a charge. The agent explains the policy, and the customer says, “Okay, that makes sense.” The transcript marks the issue resolved.

Valence registers rising irritation at the moment the policy is explained, sustained through the end of the call. The words indicate acceptance. The acoustics indicate an unresolved problem. With that signal, a human agent can address it before the call ends, and a voice agent can adjust its response instead of closing the ticket.

Order of Operations

The ear extracts acoustic features first, and the brain processes them alongside language, not after it. Most voice AI pipelines reverse that order: audio becomes text, and analysis runs on what remains. By then, the features that carry emotion are gone.

Modeling the ear means analyzing the acoustic signal directly, in parallel with the transcript. The result is a record of what a customer said and how they felt while saying it.