Undertone

A voice journal that reads tone, not just words: it scores each check-in as calm, happy, angry or sad every 2 seconds and shows where the tone changed.

Made withut-ser 3.2 tone modelSpeech emotion recognitionOn-device speech-to-text2 s windows, 0.5 s hop

Watch the film Build something like this

Facts

Status: Live
Platform: Phone app · web app
Flows: Record · Result · Sessions · Insights
Model: ut-ser 3.2 on the device: Calm · Happy · Angry · Sad per 2 s window
Input: Live recording or an audio file (WAV, MP3, M4A, up to 10 min)
Privacy: Audio encrypted and kept for 30 days; tone and transcript stay until you delete them
Year: 2026

FIG. 01 / PRODUCT FILM

Muted preview · play with sound for the full cut

Who it's for

People who talk things through out loud and want to hear how they come across: a manager recording a quick check-in before a hard call, someone keeping an evening voice diary, or anyone working on how they sound with a communication coach. Made for recording on the phone in the moment and reviewing at a desk later.

The problem

A voice note keeps what you said, not how you said it. Looking back on a hard call or a long week, you remember the words but not the moment your voice tightened, how long it stayed that way, or what you were talking about when it did.

The product

You record a check-in on your phone or at your desk, or upload an audio file, and the ut-ser 3.2 tone model on the device scores every 2-second window as calm, happy, angry or sad while the transcript is written, each phrase tagged with its tone. When you stop, the result shows the overall mix, a timeline of where the tone changed and the lines that carried it; on the desktop it also sets pitch, loudness, speaking rate and pauses against your own voice baseline, so you see what in your voice changed. Sessions keeps every check-in as a tone strip, Insights shows the patterns across weeks, and a session report can go to a coach as tone, transcript and cues, never the audio. Tone is estimated from your voice; it is not a diagnosis.

FIG. 02 / HOW IT WORKS

How it works

Step 1: Record
You speak and Undertone listens. Each bar of the waveform arrives grey and takes its tone colour one window later, the ring re-balances with every 2-second read, and the word in its centre changes only when your tone does. A live caption shows the last phrase as it lands, tagged with its tone.
Step 2: Result
When you stop, the session replays in a few seconds: the tone lanes colour in behind the playhead, the overall mix counts up and the key moments are pinned with their words. On the desktop the playhead goes back to the tense stretch — angry for 20 seconds or more at 70% confidence or higher — a bracket marks how long it lasted, and bars for pitch, loudness, speaking rate and pauses grow from your own baseline.
Step 3: Sessions
Every check-in lands in the library as a tone strip in time order, with its length, calm share, peak moment and tense count. Filter to the sessions with a tense stretch or the ones shared with your coach, and preview one — its lanes, its mix, its peak quote and related sessions — before you open it.
Step 4: Insights
Insights sets the last 7 or 30 days side by side: sessions and time recorded, the weekly tone mix against your calm average, and every tense stretch with the words heard within 30 seconds of it. It also shows the time of day you sound calmest and your voice baseline, and a weekly report can go to your coach.

FIG. 03 / SCREENS

The product, screen by screen

Undertone Record screen on a phone: a large ring reading “Calm, 72% confidence, steady for 18 s” with the Calm, Happy, Sad and Angry shares below it, a waveform coloured blue and amber by tone, the live caption of the last phrase, the recording clock at 0:42, stop and pause buttons, and a link to upload an audio file instead.

Undertone session result on a phone: the headline “Mostly calm, with a tense stretch at 1:10”, the overall mix with Calm 57%, Angry 21%, Happy 13% and Sad 9%, a four-lane timeline with the angry stretch outlined, and three key moments with their tone chips and quotes, above Replay and Export report buttons.

Undertone Sessions screen on a phone: All, Tense and Shared filter chips, a “Recording now” row with its live tone strip, and past sessions grouped by week, each with a tone strip in time order, its length, calm share and tense count.

Undertone Insights screen on a phone: a 7, 30 and 90-day switch, totals for sessions, time recorded and calm share, a weekly tone-mix chart with red dots for tense stretches, a “9 tense stretches” panel with the words heard near them, and calm share by time of day.

Web app

The same Undertone at a desk: record before a call with the live transcript beside the ring, review a session with the voice cues behind each stretch, scan the library by tone, and read the month's patterns on one screen.

Undertone web app session result: the headline with “tense stretch at 1:10” highlighted, a timeline with the waveform, four tone lanes, a “Tense · 28 s” bracket and a hatched low-confidence window; below it the transcript with a tone chip on each line, a “Why this is angry” panel with pitch, loudness, speaking rate and pauses against the speaker's baseline, and the overall mix against the last 30 days.

Undertone web app Sessions library: All, Tense moments and Shared filters over a table of sessions with when, length, a tone strip, calm share, peak moment and tense count, and a preview panel for the selected session with mini lanes, its mix, the peak moment quote and related sessions.

Undertone web app Insights page for 30 days: totals for sessions, time recorded, calm share and tense stretches, a weekly tone-mix column chart with the 30-day calm line, the tense stretches listed with the words heard near each under “7 of 9 came near vendor talk”, calm share by time of day, a calendar of sessions and the voice baseline.

FIG. 04 / DESIGN DIRECTION

Design direction

“Soft sound”, light: a warm paper ground, ink text, no cards on the ground and no shadows but one soft halo around the record button. Young Serif carries the headlines, the emotion words and the big numbers; Plus Jakarta Sans carries everything else. Each emotion has one muted colour — calm blue, happy amber, angry brick, sad violet — always shown with its label, and pills, round icon buttons and hairlines keep the screens quiet.

Palette

Ground#F5F2EC
Ink#221E2B
Calm#3A6EA5
Happy#B7791F
Angry#B23A2E
Sad#5B4B8A

Type

Young Serif: Headlines, emotion words, big numbers, clocks
Plus Jakarta Sans: Everything else: transcript, labels, buttons, charts

FIG. 05 / MOTION LANGUAGE

Motion language

“Heard, then coloured”: the motion follows how the model works. Sound arrives grey, a playhead hops one 2-second window at a time, and the tone colour lands one window behind it, because the model can only score what it has already heard. Proportions in the ring, the lanes and the charts re-balance by sliding, and words commit as whole phrases instead of typing out. Calm and unhurried, with no bounce or glow: the most dramatic thing on screen is a single word changing.

Kiko

Live

A multimodal AI assistant you can show things to: a photo, a voice note and a line of text in one ask, answered with numbered markers on your own picture.

Multimodal AISpeech-to-textOn-device face & screen blur