WORK / Timbre
Timbre
CLIENT
Timbre
YEAR
2026
SERVICES
Product Design
DURATION
1 month
Timbre is voice AI infrastructure, speech synthesis and realtime voice agents, with an observability layer underneath. Customers are engineering teams shipping voice products, support agents, IVR and long-form narration, who need audio to arrive fast and need to know why when it doesn't.
PROBLEM
Voice products fail on latency, not features
Nobody churns over a missing setting, they churn over dead air before the agent speaks. Yet latency arrives as a single averaged number, a p50 on a dashboard, that says something is slow without saying what.
- No audio to listen back to, no per-stage timing, no way to attribute delay to queue, model, or network
- A p50 on a dashboard is a verdict without evidence, so the engineer is left to guess
- Colour was chosen for mood, so every chart and waveform carried hues with no meaning
- Brand and data drifted apart the moment either one changed, with no legend to read them by
APPROACH
Make the palette the legend
The colour system is a single cool hue family running ten steps from abyssal navy to pale aqua, where position on the ramp encodes intensity. The same tokens that carry the brand render every waveform, spectrogram and chart in the product. Colour stops being ornament and becomes the key to the data. On that foundation, time to first audio is decomposed into five measured stages per response rather than averaged away. Every voice carries a signature print generated from its own name, so fourteen voices stay distinguishable at a glance and each one is tracked separately. The mobile app was scoped to the two things a phone does better than a desktop, alerting and listening, rather than ported from the dashboard.
IMPACT
From a number, to a cause
76%
Of a latency breach traced to one stage, in a single view
100%
Of product visualisation driven by the brand's own tokens
0%
Variance between figures across eight surfaces
The measure of a system like this is what it lets you conclude. A session flagged at 604 ms means nothing on its own. Broken into five stages it resolves at once, queue 38, normalisation 22, inference 421, vocoder 96, egress 27. Against a 148 ms median whose stages sum to 12, 18, 74, 33 and 11, model inference is running at 5.7 times typical and accounts for 76 per cent of the excess. Because every bar, cell and fill resolves through the same ramp, retuning one primitive moves the mark, the marketing site and every visualisation together. The palette was rebuilt entirely mid-project, from a warm spectrum to a cool one, and not a single chart needed redrawing.
