Three rules, and a miss is a result
- The pass mark comes first. We write down the criteria and the decision rule before we measure anything.
- New data. The recordings we test on are not the ones we built the measure on.
- A miss is reported. A measure that misses its mark keeps a lower label, or is not offered, and we say so. Several of the areas below are there because a test said no.
Area by area
Loudness, true peak and levels
Verified against a standard or a referenceTested: Integrated, short-term and momentary loudness, loudness range and true peak, against the EBU Tech 3341 and 3342 test signals, on every build of this site.
What it means for you: Use it for delivery checks. For a formal broadcast audit a certified meter is still the reference.
Spectrum, stereo, clipping and file facts
Verified on known signalsTested: Known tones, noise, clipped signals and files, where the answer follows from the definition.
What it means for you: The calculation is right. Each measure says on /lenses what it was not tested on.
Room acoustics: reverberation from a clap
Verified against a standard or a referenceTested: Reverberation time per octave band on hundreds of real impulse responses, against our RoomLab app and an independent implementation.
What it means for you: The clap in the tests is simulated; real phones and real claps are the next test.
Speech, pauses and signal-to-noise
IndicativeTested: Read speech from new speakers in many noise conditions, with the pass mark written down first.
What it means for you: Good on clean to moderately noisy speech, with a margin published per measure. Below about 10 dB signal-to-noise it reads optimistic.
What kind of audio this is
IndicativeTested: Fragments of speech, music and noise, and speech over music, on recordings the model never saw.
What it means for you: Speech and music are recognised well by the newer model (opt-in). Steady noise and speech over music are weaker, and we say so next to the result.
Background scene
IndicativeTested: Nine real noise environments at several levels.
What it means for you: Two measures sort crowd noise from machine noise as groups of environments. A hint for a collection of files, not a verdict on one file.
Signal chain: lost packets
Verified on known signalsTested: Gaps of digital silence cut into speech, with a known place and length.
What it means for you: A lower bound for real calls: a loss that the codec concealed leaves no gap to find.
Echo and slapback
Under testTested: Not published yet.
What it means for you: It is not offered through the API or the MCP server, and it has not been tested on real calls.
Pitch and voice quality
Verified on known signalsTested: Pitch and harmonics-to-noise ratio on constructed voices and on real read speech.
What it means for you: Jitter and shimmer are not offered: they narrowly missed the pass mark we set.
Predicting speech-recognition errors
Not possible from audio aloneTested: Whether audio measures predict which words a speech recogniser gets wrong, in a test with the pass mark set first, two recognisers and invented but realistic reports.
What it means for you: They did not. The recogniser’s own confidence did, in part. We do not claim that our measures predict transcription errors.
What the levels mean: Verified against a standard or a reference. Checked against a published standard or an independent implementation, with the source named. Verified on known signals. Checked on audio where the right answer is known because we made it. It shows the calculation is right, not that every real recording reads the same way. Indicative. A useful pointer with a stated margin. It did not meet the pass mark we set for a measurement, so use it as a direction. Under test. Built and tested, but not offered, or not yet tested on real recordings. Not possible from audio alone. We tested whether it can be done and found that it cannot be claimed.
The reports, or a test on your audio
The reports
For every area there is a written report: the question, the recordings, the pass mark, the result, and what did not work. We share them on request with customers, partners and researchers.
A test on your recordings
We run our measures on a sample of your audio and tell you how far each number can be trusted for that material.
A question we have not tested
If you need to know whether something can be measured reliably in your audio, ask. We will tell you what we know and what a test would take.