/docs
Documentation: the API, MCP, and every lab.
Written like a working engineer wrote them, because one did: quickstarts for the hosted analysis API and the MCP server, plus practical references on DSP, Android realtime audio, accessibility patterns, voice QA, and the engineering choices behind each lab.
Start here
Do
What’s new 13 October 2026 Three new profiles: room (reverberation from a clap), chain (lost packets) and pitch (where a voice sits and how clean it is).
Building with an LLM? Every page is also plain markdown: /api.md · /llms.txt
MixLab 3
- IntroWhat integrated LUFS actually tells youIntegrated LUFS is the loudness number you should care about, and the one most "loudness" tools get subtly wrong.
- PracticalReading crest factor without fooling yourselfCrest factor is the easiest dynamics metric to compute and the easiest to misread.
- DeepMid/Side decomposition in WebAudioSum and difference signals are the right way to think about stereo. Here is the WebAudio API code to do it.
VoiceLab 8
- IntroVoice AI: ten questions, answeredWhat AudioLab measures for a voice agent, what it does not, where your audio goes, how a verdict is made and who sets the limits, how to compare versions and automate the checks, and what is free.
- IntroWhy filler density matters more than filler countCounting "ums" is the wrong measurement. Density tells you whether they are a problem.
- PracticalA pragmatic loudness target for podcastsThe right loudness target for spoken-word podcasts is not the same as music. Here are the numbers and why.
- PracticalRead the audio before you transcribeOne call before the model: what the scene fields tell you about a file, and what they do not, with the numbers from our own transcription test.
- DeepWhat envelope decay after speech actually measuresWe built a cheap reverb proxy from the decay after each phrase, then tested it against 24 measured rooms. It does not measure reverb. Here are the numbers and what we changed.
- PracticalMeasure how long a room rings with one clapThe room profile reads the reverberation time per octave band from one clap, a balloon pop, an impulse response or a sweep. How to record it, how to read the answer, and how far the numbers are tested.
- PracticalPitch and voice quality: what the profile gives and what it holds backThe pitch profile reads where a voice sits, how much it moves, how voiced the recording is and how clean the voice is. The test numbers, the limits, and why jitter and shimmer are not offered yet.
- PracticalFind lost packets in a recordingThe chain profile counts the gaps of digital silence that a lost packet leaves in the middle of speech. How it decides, the test numbers, what it cannot see, and why the echo and slapback measure is not offered.
HearLab 2
- PracticalLive Caption on Android: what it does and doesn't doGoogle’s Live Caption is the most-used accessibility feature in the world. Here is how it actually works and what it doesn’t cover.
- DeepHearing-aid routing through Android Audio FrameworkHow AAudio, ASHA, and LE Audio actually deliver audio to hearing devices on modern Android.
SignalLab 3
- IntroA useful tag schema for audio archivesWhat to put in your audio metadata so downstream tools can actually use it.
- PracticalStreaming clipping detection at scaleHow to detect clipping in audio at ingest time without decoding the whole file.
- DeepSpeaker turn segmentation: the pragmatic stackDiarisation is a hard ML problem. Speaker turn segmentation, the cheap version, is solved enough to use everywhere.
CueLab 3
- IntroA reusable pre-show checklistThe checklist that survives the moment when your mind goes blank thirty seconds before going live.
- PracticalOBS audio routing without surprisesOBS audio routing is more powerful than the default UI suggests. Here is the mental model that makes it predictable.
- DeepDesigning cue-list data that survives a re-orgA cue list is a data model first and a UI second. Treat it like one.
SkillLab 3
- IntroDesigning tiered audio challengesBronze / Silver / Gold tiers exist for a reason. Here is the structure that produces real skill gain.
- PracticalFeedback that lands: patterns from real coachesGeneric feedback is worse than no feedback. Here are the patterns that real coaches use.
- DeepProgression loops without slot-machine mechanicsHow to keep people coming back without weaponising dopamine.
General 6
- PracticalCompare voice-agent and TTS vendors, and check conversational audioThree API calls, ten numbers, and how far to trust each one: a guide for QA teams comparing voice-agent, text-to-speech or voice-isolation vendors.
- PracticalCheck audio quality in CI: fail a build when a file breaks your rulesA short Python script and a GitHub workflow that measure each audio file with the AudioLab API, compare the numbers with limits you set, and fail the build with a message that says why.
- PracticalAudit a dataset of conversations: where the audio really comes fromMeasure a folder of call, voice-agent or text-to-speech recordings for a lossy source, upsampling, splices, mono in disguise, clipping and bit depth, 20 files at a time, and know when the audio is uploaded and when it is not.
- IntroRead a response: what each field means and how far to trust itA real answer from the API, field by field: what it says, its status and margin, and what you may and may not conclude from it.
- IntroWhat AudioLab does not doThe limits of the analysis, stated plainly: it does not transcribe, predict recognition errors, score naturalness, identify speakers or measure latency, and its background hints are guesses.
- IntroChanges and deprecation: what changes, when, and how you find outThe promise behind /v1, what counts as a change that needs notice, how we announce it, and what is deprecated today.