AudioLab Voice
Quality control for the audio your voice agent hears and sends
157 measures, each with a status and a margin. In your browser, your audio is never uploaded.
Want to know how we measure first? The path from microphone to transcript, step by step →
What are you checking?
- Calls and conversations Available
Check a call
Level, clipping, noise, gaps and the band of a recording, one at a time in the browser or twenty at a time over the API. Check a call →
- Datasets Available
Open the dataset view
Every file judged by the same limits: pass, check or fail, the rules that fail most, the files to look at first. Open the dataset view →
- Text-to-speech output Available
Judge your output
Loudness and true peak against your target, clipping, noise floor, gaps, bandwidth, pitch and voice quality, with a verdict traced to a rule. Judge your output →
- Recogniser stress test In preparation
Tell us you want it
A fixed set in known conditions through your own recogniser or agent, and the change in word error rate for each condition. No date yet. Tell us you want it →
- Live stream Demo
Try the live monitor
A monitor that names events as they happen in a stream: a gap, clipping, a level drop, dead air, hum. Running as a demo in your browser. Try the live monitor →
- Your own limits Available
Set your limits
Set limits per measure, see which rule a verdict rests on, and take them into your pipeline as a gate file. Set your limits →
Free: 300 minutes a month and everything in the browser. Packages: 5,000, 25,000 or 100,000 minutes for €49, €149 or €399, one-time, valid 90 days. All prices →
Three moves, no account for the first
01
Measure in your browser
Pick a recording and read every number with its status and margin. Your audio is never uploaded.
02
Automate the same checks
The same measures over the API and MCP, with a free key: 1,000 calls a month, up to 20 recordings in one batch call, and a quality gate for your build.
03
Compare, and test the recogniser
Two versions of the same utterance against the margin of each measure, a set of files judged by one set of limits, and a stress test of your recogniser in known conditions, in preparation.
157 measures, and how sure we are of each 94 validated · 56 proxy · 5 beta · 1 withdrawn · 1 planned
- ● Validated
- 94
- Tested against a standard or a known value, with the source named.
- ▲ Proxy
- 56
- Approximates something. Only the direction is tested, or the thresholds were chosen by hand.
- ◆ Beta
- 5
- No ground-truth test yet. It may be right; it is not shown to be. It can change.
- ✖ Withdrawn
- 1
- Failed its test. Kept in the changelog, no longer shown as a result.
- ○ Planned
- 1
- In the registry, not built.
One square per measure, coloured and shaped by its status: ● validated against a standard or a known value, ▲ a proxy that approximates something, ◆ beta without a ground-truth test yet, ✖ withdrawn because it failed its test. Click a square for the measure; the atlas puts all of them on the path from the microphone to the recogniser.
Voice or audio, and how much a month
The browser labs and checks are free and use no calls. For the API and MCP, say what you work with and how much, and see which tier carries it; the arithmetic is on the card.
Every number carries its own doubt
- A status is a promise about evidence, not about quality. Validated means tested against a standard or a known value, with the source named. A proxy is information, never a pass or fail. Beta is shown nowhere in the voice checks.
- A margin says how far the number can be off. A verdict closer to its limit than the margin says "check", not "fail".
- We publish the tests that failed, including a reverberation score we withdrew, on where we were wrong. A measure does not predict how many words a recogniser gets wrong, and none of our pages says it does.
- Your audio. The checks in the browser measure on your device and upload nothing. The API receives the audio at the EU-hosted service and keeps only numbers; see the path page and /privacy.
- The whole path, step by step Ten questions, answered How we test