Start listening

REC — Voice agent observability

Your voice agents are talking.
Are you listening?

Timbre is the studio console for AI voice agents. Every call scored for sentiment, every millisecond of latency accounted for, every transcript searchable — before your customers hang up on a bad take.

Start listening — free Open the console

$ npm i @timbre/sdk — first 1,000 minutes on us

In production at ParceloDuskbankNorthbeam AirOndo HealthKettle & CoFerryline
512M+minutes of agent audio analyzed
1.2s → 380msmedian turn latency after tuning with timbre
31%fewer human escalations in the first month

The signal chain

Every call, mixed down to what matters

Logs tell you the call happened. Timbre tells you how it sounded — where the model hesitated, where the customer's patience ran out, and which prompt change fixed it.

Transcripts

Transcripts that know what they mean

Every utterance is tagged with intent, entities, and outcome. Search "angry + refund + interrupted" across a million calls and get an answer in under a second.

Sentiment

Mood, second by second

A sentiment track runs under every call like a waveform. See the exact sentence where a save became a churn.

Latency

Find the slow stage

A waterfall for every turn: ASR, LLM, TTS, network. When p95 drifts, you'll know which stage — and which deploy — did it.

Alerts

It pages you before the CSAT survey does

Set thresholds on anything Timbre measures — sentiment dips, latency spikes, silence gaps, repeated intents. Alerts land in Slack or PagerDuty with the offending audio attached.

Wire it up

From silent to signal in an afternoon

Patch in the SDK

One middleware wraps your voice pipeline — Twilio, LiveKit, Vapi, Pipecat, or raw WebRTC. Audio is fingerprinted and streamed, never stored raw unless you say so.

// three lines, honestly
import { timbre } from "@timbre/sdk";
pipeline.use(timbre({ agent: "checkout-v3" }));

Timbre scores every second

Sentiment, intent, interruptions, silence, latency per stage — computed live, indexed instantly. The console lights up with your first call.

avg time to first insight: 4 min 12 s
overhead added to your calls: 0 ms — we tap the stream, we're not in it.

Tune, deploy, hear the difference

A/B prompts and voices against real metrics, not vibes. Timbre attributes every sentiment and latency shift to the exact deploy that caused it.

prompt v41 vs v42: csat +0.4, e2e latency −210 ms, escalations −1.3 pts → ship v42

Pricing

Pay for minutes, not seats

Everyone on your team gets the console. You're billed only for the audio Timbre analyzes.

Demo tape

$0

For prototypes finding their voice

  • 1,000 analyzed minutes / month
  • Live console & transcripts
  • Sentiment & latency tracks
  • 7-day retention
Start free
In heaviest rotation

Studio

$149 /mo + $0.006/min

For voice agents in production

  • 25,000 minutes included
  • Alerts to Slack & PagerDuty
  • A/B prompt & voice experiments
  • Deploy-level attribution
  • 90-day retention, full audio
Start 14-day trial

Label deal

Custom

For fleets of agents at scale

  • Unlimited minutes, volume pricing
  • SSO, RBAC & audit logs
  • In-region processing (US/EU/APAC)
  • Custom scoring models
Talk to an engineer

Put your ear back on the line.

Your agents handled ten thousand calls this week. Timbre can tell you about every one of them — starting four minutes from now.

Start listening — free no card · 1,000 minutes free · cancel anytime