What AudioLab does not do

AudioLab measures the signal in a recording. This page lists what it does not do, so that nobody builds a pass or fail rule on a number that cannot carry it. Each point says what to use instead.

It does not read the words

There is no transcription, no language detection and no check that the text was spoken as written. Use a speech recogniser for the words and put the AudioLab numbers beside its result.

It does not predict transcription errors

We tested this on 10 October 2026: 624 recordings of invented Dutch citizen reports, read by three synthetic voices in 13 conditions, and 1,248 transcriptions by two Whisper models. The signal-to-noise measures did not predict whether the street, number or town came back wrong: the AUC was 0.52 to 0.55, where 0.50 is a coin flip. The intelligibility label (sceneIntelligibility) flagged 58 percent of all recordings as not clear, and of the flagged ones 25 to 47 percent had a real location error.

So do not refuse a file on one of these numbers, and do not say that a better signal-to-noise ratio means a better transcript. The test used synthetic voices, Whisper only, a simulated telephone band and Dutch text; real speakers, other recognisers and other languages are untested. The numbers and the limits are on Read the audio before you transcribe.

It does not say whether a voice sounds natural

None of the measures on /lenses scores naturalness, human-likeness or how robotic a voice is. Speaking rate, sibilance risk, clipping, the noise floor and the bandwidth describe the signal, and they find technical defects. Whether a voice sounds right needs listeners.

It does not tell voices apart

There is no speaker identification, no counting of speakers and no diarisation. A few measures hint that something with a voice-like cadence sits in the pauses (sceneGapHarmonicShare, and sceneBackground reading program). They say that it is there, not whose voice it is.

It does not measure latency, interruptions or turn-taking

It analyses one recording. How long the agent took to answer, whether it was cut off and who spoke when across a call need timestamps from your own system, not from an audio file. Inside one recording, longSilences counts the pauses over 1.5 seconds.

Its background hints are guesses

sceneBackground (quiet, stationary, program or unsteady), the other scene fields, activeKind and contentTypeGuess are proxies read from the pauses and from the frames. The background word was tested on synthetic scenes only, not on real recordings. On the clean synthetic voice in Read a response, sceneBackground says “unsteady” and activeKind says “music”. Use them to know where to look, not to decide.

It does not set your limits

There is no standard pass level for conversational audio. Our own presets (broadcast, streaming) exist for delivery, and a voice agent is not delivery. The starting ranges in Compare voice-agent and TTS vendors show what we measured, not what is good. Set your limits from your own good recordings.

What it was tested on

• Read speech. The speech detector behind the speech ratio, the pauses and the scene fields was measured on 40 read English recordings (LibriSpeech) in 17 noise conditions. Spontaneous conversation has not been tested.
• Known noise at known levels. The margin of the signal-to-noise ratio comes from the same 40 recordings with real background noise added at known ratios: steady noise (kitchen, living room, hallway) at a true ratio of 10 to 30 dB, and babble at 20 dB. Below 10 dB it reads too high. Other kinds of noise, such as traffic, were not tested.
• Not tested. Other languages, phone-band audio, music or television behind the voice, and reverberant rooms, for the measures that lean on the speech detector.

Every measure says what it was tested against and what it was not on /lenses. A status of Validated means tested against that, under those conditions.

When you need one of these

Write to support@audiolab.tools with what you are trying to decide. If a measure that would help is not on the list, that is useful to know, and we will say plainly if it is something we cannot do.