An independent audit of your audio, with numbers you can re-run
We measure your recordings and tell you, with rule and number, what we find and which problems are
worth fixing. Clipping, dropouts, loudness, noise floor and bandwidth are part of the report; how recognition holds up on names,
addresses and numbers is a project on request. This is a measurement report, not a certification.
In the browser your audio is not uploaded. Through the API it is processed in the request and an upload is gone within the hour; it is
not used to train models, and not sold or shared. Where your audio goes
Start with a free scan
Send 5 to 10 files your agent really sent, ideally after the phone line or codec. Within two working days you get five lines in
plain words: what the check found, the rule and number behind each finding, and what it cannot see. No obligation. A few teams at a time.
✓Up to 40 files from one pipeline, measured in my own browser, so your audio is never uploaded to AudioLab.
✓An overview: pass, check and fail, the rules that fail most, and the files to listen to first.
✓A plain-language PDF report that also says what the measures cannot see.
✓A 30-minute call.
✓Delivered 2 working days after I receive the files.
Full
€790 excl. VAT
✓Everything in Compact, for up to 100 files, or for two sets you want to compare, such as before and after a codec change.
✓A CSV with the numbers per file, and a provenance check: sample rate, bandwidth, clipping, channels and bit depth. For phone and speech-codec audio the check is limited to those.
✓Your limits as a gate file for your CI, with a working example.
✓A 45-minute call, and one retest after your fix.
✓Delivered 3 working days after I receive the files.
Team pilot
On request
✓For product teams that want the check inside their build: the Full audit, and the gate file ready to merge with your limits.
✓A 90-minute team session.
✓Then a monthly re-check of a fresh sample with a short written note, and email support with an answer within one working day.
✓The first quarter is fixed; after that it continues month by month. Ask for the scope and the price.
If you start with Compact and then want the gate and the CSV, you pay the difference within 30 days, on the same files.
Not a conversation test: no turn-taking, latency, naturalness, pronunciation or the words that were said.
No prediction of recognition errors. In our own test the SNR measure did not predict them: what we tested.
No monitoring around the clock. The live monitor is a demo.
No certification or legal opinion. "Audit" here means a measurement report.
Measured on read speech and constructed signals, not on real conversations. The start limits come from 48 synthetic recordings (16 Dutch sentences read by 3 synthetic voices) and are
not a standard: your limits are yours. Speech share and pauses are indicative. Of the measures, 97 of 157 are
Validated today; the rest say on /lenses how far to trust them.
Your audio stays on your machine. You run the dataset view yourself in
your browser, and we look at the result together on a call with screen sharing. The dataset view has no export. In our test one
Chrome tab took 200 files of 30 seconds without slowing down; it holds about 10 MB per file for mono and 20 MB for 48 kHz stereo,
so a laptop with 8 GB takes about 100 short files (an estimate; Firefox, longer files and a laptop with 8 GB were not measured).
You send me files. I measure them in my own browser, use them for nothing else and delete
them after delivery. I sign a data processing agreement or an NDA first.
Through the API. For larger sets: the files go to the EU-hosted service, only numbers come
back, and the audio is processed straight away or kept for at most an hour. Our data processing addendum
applies; where your audio goes has the details. Start with files you are free to share.
On your machine, with AudioLab Local (beta). After an NDA we run the measurements together on your own machine: a folder of recordings in, a verdict per file and a report out, and your audio is not uploaded. It reads phone recordings at their own sample rate. More about AudioLab Local.
Never send a key or password in a message; you make your own free key.
What I need from you
5 to 10 files for the free scan, 40 or 100 for the packages, in wav, mp3, flac, m4a, ogg, opus, aac or aif. Say where in your pipeline
they were taken, ideally after the phone line or codec. Your limits if you have them; otherwise we start from the EBU R128 preset or
the start limits and tune. A contact for the call.
You get the report, the numbers and, in Full, the gate file and the working example.
We talk it through, and in Full I retest after your fix.
Before you ask
Use Chrome or Firefox for the browser checks; Safari and iPhone are not tested, and the capacity figures above were measured in Chrome only.
Until 15 January 2027 the API serves the first models: send "speechDetector": "k2" and "contentModel": "v2" to get the same numbers as the browser.
The browser limit is 50 MB per file; the API decodes at most 170 MB of audio per file (about 7.7 minutes of 48 kHz stereo or 46 minutes of 16 kHz mono).
The gate file has no field for dropouts; the download says so.
In the browser a 16 kHz recording reads a lower bandwidth than the file holds (we measured 5.8 to 7.8 kHz on 16 kHz speech), so it can fail the start limit of 7,900 Hz while the band is full. The report says so for phone and 16 kHz audio.
What we already know is weaker is listed on try to break it.
Questions
Is my audio uploaded?
Not in the browser route. In the API route it goes to the EU-hosted service, see above.
Why no single score?
An average hides a clipped file. Every verdict traces to a rule, a number and a limit.
What if you find a wrong number?
It goes on our public where we were wrong page, with your name if you want it.
Do you offer a subscription?
After an audit you can keep the gate running on the Studio plan or talk to us about Enterprise (from €249 a month, on request).