/glossary

Terms, plainly defined.

A working glossary of audio, AI, DSP, accessibility, and live-audio terms used across the labs. Updated whenever a doc surfaces a new term that needs one.

82 terms

AAudio Engineering: Android’s native low-latency audio API, introduced in Android 8.0 and matured significantly in 12+. The recommended path for realtime audio on modern Android.
ABX test SkillLab: A double-blind listening test: the subject is presented with reference A, reference B, and unknown X, and must identify whether X = A or X = B. The gold standard for proving an audible difference exists.
AGC VoiceLab: Automatic Gain Control: adaptive amplification that keeps signal levels consistent without operator intervention. Standard in conferencing apps; used carefully in podcast post-production.
Aliasing Engineering: The spurious low-frequency artefacts produced when a signal contains frequencies above half the sample rate (the Nyquist frequency). Prevented by anti-aliasing filters before sampling and by oversampling in nonlinear DSP.
ASHA HearLab: Audio Streaming for Hearing Aids: the Bluetooth profile used by Android to stream low-latency audio directly to compatible hearing aids.
ASR VoiceLab: Automatic Speech Recognition: the speech-to-text layer. Whisper, Deepgram, AWS Transcribe, Google STT. Mostly a commodity in 2026 at high-90s % accuracy on clean material.
Audiogram HearLab: A graph of hearing threshold (dB HL) versus frequency (Hz), per ear. The standard diagnostic output of an audiology test. Drives the prescription of any modern hearing aid.
AudioWorklet MixLab: The browser API for running audio processing code on a dedicated thread, separate from the main thread. The foundation for any serious WebAudio analysis or synthesis.
Auracast HearLab: Bluetooth LE Audio broadcast mode. A single source streams audio that any nearby compatible device (including hearing aids) can tune to without pairing. The basis for next-generation assisted-listening systems.
Bit depth General: The number of bits per sample. 16-bit covers most consumer use; 24-bit is the production standard; 32-bit float is the working format inside DAWs and analysis tools.
Brickwall limiter MixLab: A look-ahead peak limiter with an absolute ceiling: set the threshold and no sample passes above it. The last stage of nearly every loudness-normalised master.
BS.1770-4 MixLab: The ITU recommendation defining how to measure loudness in audio. The basis for LUFS metering, EBU R128, and streaming-platform normalisation targets.
Buffer size Engineering: The number of audio samples processed per callback. Smaller = lower latency but more CPU per second; larger = higher latency but more efficient. 128 or 256 samples is the modern realtime sweet spot.
Codec General: A coder-decoder pair that compresses and decompresses audio. Lossy: MP3, AAC, Opus, LC3. Lossless: FLAC, ALAC, WAV/PCM (uncompressed). The choice of codec sets the floor for what any downstream tool can achieve.
Crest factor MixLab: The ratio between peak and RMS levels in a signal, expressed in dB. Higher numbers indicate more dynamic range. See the crest factor doc.
Critical listening SkillLab: Trained, focused listening for specific audio attributes: frequency balance, dynamics, distortion, stereo width. The skill SkillLab drills systematically.
Cue list CueLab: An ordered sequence of show-runtime triggers: audio, lighting, video, scenes. The data structure underneath every live-event production tool. See CueLab’s cuelist schema doc.
DC offset General: A constant shift of the waveform away from the zero line. Inaudible by itself, but it wastes headroom and can cause clicks at edits. A high-pass filter around 20 Hz removes it.
De-esser VoiceLab: A frequency-selective compressor that targets the 5–9 kHz sibilance band. Reduces harsh "s" and "t" sounds in vocal recordings without dulling the overall top end.
Diarisation SignalLab: Speaker turn segmentation: identifying who spoke when in an audio file. AudioLab does not do this. The SignalLab indexer segments by signal (voice, music, noise, silence), not by speaker. For diarisation, reach for pyannote or NeMo.
Dither MixLab: Low-level noise added before bit-depth reduction to randomise quantisation error. Required when converting from 24/32-bit to 16-bit for distribution.
Ducking VoiceLab: Reducing the level of one signal whenever a higher-priority signal is present. The classic example: music ducking under voice in a podcast or radio show.
Dynamic EQ MixLab: An EQ band whose gain reacts to level. Conceptually a single-band compressor with a parametric frequency control. Less destructive than multiband for surgical tonal balancing.
Embedding SignalLab: A learned dense vector representation of audio (or a chunk of audio). The core of modern audio retrieval, similarity search, and zero-shot classification. AudioLab signal indexing uses embeddings under the hood.
FFT MixLab: Fast Fourier Transform: the algorithm at the heart of every frequency-domain operation in audio software, from spectrum analysers to convolution reverbs.
Formant VoiceLab: A resonant frequency band in the human vocal tract. Formant patterns distinguish vowels and reveal speaker identity. Preserved or shifted intentionally by pitch-correction tools.
Hann window MixLab: A common windowing function applied before FFT to reduce spectral leakage. The default windowing choice for most analysis-oriented audio code.
Headroom MixLab: The dB margin between the loudest sample in a mix and 0 dBFS (or true peak). Streaming masters typically target –1 dBTP / –14 LUFS; broadcast targets are more conservative. See the methodology page for AudioLab’s defaults.
HRTF HearLab: Head-Related Transfer Function: the frequency-domain filter describing how a sound from a specific direction is altered by the listener’s head, torso, and outer ear. The basis for binaural spatial audio.
Inference SignalLab: Running a trained model to produce outputs. Distinct from training. The phase that needs to be fast, predictable, and small enough to fit on-device. The user-visible part of any audio AI product.
Integrated LUFS MixLab: A perceptually-weighted, time-gated measure of loudness across an entire piece of audio. The number streaming platforms use to normalise playback.
Intersample peak MixLab: A peak that falls between two samples, so a meter that only reads samples misses it. After playback conversion or MP3/AAC encoding it can exceed full scale and clip. True-peak meters (dBTP) estimate it.
JND SkillLab: Just Noticeable Difference. The smallest change in a stimulus that a listener can reliably detect. Drives the calibration of every audio-perception threshold in SkillLab’s drills.
JUCE Engineering: Cross-platform C++ audio framework. The path of least resistance for shipping the same DSP code on Android, iOS, macOS, Windows, and as VST/AU/AAX plugins.
K-weighting MixLab: The pre-filter chain (high-shelf at ~1.5 kHz + high-pass at ~38 Hz) defined in BS.1770-4. Applied before RMS energy summation so the resulting LUFS value tracks perceived loudness rather than raw signal energy.
LC3 HearLab: Low Complexity Communication Codec. The default codec for Bluetooth LE Audio. Materially better than SBC and competitive with AAC at lower bitrates.
LE Audio HearLab: The Bluetooth audio specification introduced in 5.2, including LC3 codec and multi-stream profiles. Material improvement for hearing devices and low-power audio.
Lissajous figure MixLab: A plot of one signal against another, producing a curve whose shape reveals their phase relationship. The classical visualisation behind every "phase scope" in pro audio.
Loudness normalization MixLab: Playback that raises or lowers every track toward a common loudness, so levels match from song to song. Spotify and YouTube use about −14 LUFS as the reference. It changes playback only, never the file, and a louder master is simply turned down rather than marked as bad.
LRA (Loudness Range) MixLab: EBU 3342 metric expressing the difference between the loudest and quietest sections of a piece of audio. Distinct from peak-to-RMS dynamics; closer to perceived "how much the music breathes".
Mel spectrogram SignalLab: A spectrogram on the mel frequency scale, with perceptually-weighted bins that more closely match human pitch perception than linear frequency. The standard input representation for audio neural networks.
MFCC SignalLab: Mel-Frequency Cepstral Coefficients. The classical compact representation of timbral content: 13–40 coefficients summarising the spectral envelope. Still useful as features even in the deep-learning era.
Mid/Side MixLab: A stereo representation where Mid = (L+R)/2 (centre content) and Side = (L−R)/2 (stereo difference). The basis for measuring true stereo width.
Mix-minus CueLab: A feed sent to a remote guest containing the full programme audio except their own voice, which prevents the “talking-into-a-tunnel” feedback experience. Standard in broadcast and remote-podcast routing.
Momentary LUFS MixLab: A 400 ms sliding-window loudness measurement, used to catch transient loud passages. Calibrated against the same K-weighting filter as integrated LUFS but without gating.
Multiband compression MixLab: Splitting the signal into 2–6 frequency bands and compressing each independently. The standard tool for taming a single problematic band (boomy lows, harsh upper-mids) without dulling the whole mix.
MUSHRA SignalLab: MUlti Stimulus test with Hidden Reference and Anchor. The standard listening-test methodology for evaluating perceived audio quality at intermediate-to-high quality levels.
N-1 feed CueLab: Another name for mix-minus. Literally: a mix of all N sources minus the one going to that destination. Built natively into proper broadcast consoles; emulated via routing matrices in software.
Neural vocoder VoiceLab: A model that converts an intermediate audio representation (typically a mel spectrogram) into a waveform. WaveNet, HiFi-GAN, BigVGAN. The “last mile” of any modern TTS pipeline.
Noise floor General: The level of the quietest parts of a recording, usually background noise or hiss. The lower it sits, the cleaner the recording. In music, quiet passages are often the music itself, not noise.
Nyquist frequency Engineering: Half the sample rate: the theoretical upper bound of representable frequencies in a digital signal. 22.05 kHz at 44.1 kHz; 24 kHz at 48 kHz. The reason mastering chains often work at 96 kHz or higher.
Oboe Engineering: Google’s C++ wrapper around AAudio (with OpenSL ES fallback for legacy devices). The recommended starting point for Android audio code in 2026.
OfflineAudioContext MixLab: WebAudio API variant for offline, faster-than-realtime audio processing. The right choice for batch analysis and rendering. Powers the AudioLab sample-synthesis pipeline.
Onset detection SignalLab: Identifying the start times of musical events: drum hits, note attacks, percussive transients. Foundational for beat tracking, tempo estimation, and segment-level audio analysis.
Pearson correlation MixLab: The statistical measure of linear correlation between two signals, returning a value from −1 to +1. Used in MixLab to compute stereo correlation between L and R channels.
Phantom power VoiceLab: +48 V DC supplied to condenser microphones via the same XLR cable that carries audio. Required to operate the mic capsule; supplied by audio interfaces and mixers.
Phase / polarity Engineering: Polarity is binary (signal inverted or not). Phase is continuous and frequency-dependent. Conflating the two is the most common audio-engineering vocabulary error, and the cause of many comb-filter mysteries.
Phoneme VoiceLab: The smallest unit of speech sound that distinguishes one word from another in a given language. English has roughly 44; phoneme-level alignment underpins every modern TTS and ASR system.
Plosive VoiceLab: Low-frequency burst from "p" and "b" consonants hitting the microphone. Mitigated by pop filters in capture and high-pass filtering or transient editing in post.
PLR (Peak to Loudness Ratio) MixLab: The gap in dB between a track’s loudest true peak and its average loudness (integrated LUFS). A low PLR, under about 8 dB, means heavy limiting. 8 to 12 dB is typical of modern masters, and higher keeps more dynamics.
Pre-show checklist CueLab: The reusable set of pre-broadcast / pre-event verifications that catches preventable failures. Hardware-routing-levels-stream-backup, in that order. See the pre-show checklist doc.
Prosody VoiceLab: The rhythm, stress, and intonation of speech. The hardest thing for synthetic voices to get right, and the most common giveaway when they don’t. Drives intelligibility and perceived emotion.
Round-trip latency Engineering: The total time from sound capture to sound output through a processing chain. The headline number for any realtime audio app; under 15 ms is the threshold for monitoring-while-tracking.
RT60 VoiceLab: The reverberation time required for an impulse to decay 60 dB. The standard metric for room acoustics. Practical voice-QA tools estimate it implicitly via envelope-decay analysis.
Sample rate General: The number of samples per second. 44.1 kHz for CD; 48 kHz for broadcast and most modern production; 96 kHz / 192 kHz for high-resolution mastering.
Short-term LUFS MixLab: A 3-second sliding-window loudness measurement. The middle layer between momentary (400 ms) and integrated (whole-program). Drives the “LRA” range calculation in EBU 3342.
Sibilance VoiceLab: High-frequency consonant energy from "s", "sh", "t", and "z" sounds. Concentrated around 5–9 kHz. A common source of listener fatigue when over-emphasised.
Sidechain MixLab: An external control signal that triggers a dynamics processor (typically a compressor). The technique behind pumping kick-drum mixes, voice-ducked music beds, and modern de-essing.
SNR General: Signal-to-Noise Ratio. The dB margin between desired signal and background noise. A basic measure of recording quality and a rough guide to how easily a listener can follow the voice. It does not reliably predict how well a speech recognition model will do: in our own test of 1,248 transcriptions it did not.
Source separation SignalLab: Splitting a mixed audio signal back into individual stems (vocals, drums, bass, other). Demucs, Spleeter, Open-Unmix. State of the art in 2026 produces clean isolation on most pop music.
Speaker verification VoiceLab: Confirming a claimed speaker identity from a voice sample. The defensive counterpart to voice cloning. Modern systems use d-vector or x-vector embeddings compared via cosine similarity.
Spectral centroid MixLab: The "centre of mass" of the spectrum: a weighted average frequency that corresponds roughly to perceived brightness.
Spectral flatness MixLab: How noise-like a sound is. Tonal sounds such as notes score low, while hiss and noise score high.
Speech-in-noise HearLab: A class of metrics and tests (SIN, QuickSIN, HINT) measuring speech intelligibility against varying noise floors. The real-world predictor of hearing-aid benefit, more so than pure-tone audiometry.
Streaks SkillLab: Practice-consistency markers used in SkillLab. Designed to survive a missed day to avoid slot-machine progression dynamics. See progression design.
Tag schema SignalLab: The controlled vocabulary used to describe audio content for downstream retrieval. SignalLab’s schema covers content type, brightness, dynamics, and issue flags. See the tag schema doc.
True peak MixLab: The actual analog peak of a reconstructed signal, which can exceed the sample-domain peak due to inter-sample peaks. Measured via oversampling.
TTS (Text-to-Speech) VoiceLab: Generating speech audio from text. State of the art in 2026 is neural: typically a text encoder + acoustic model + vocoder pipeline, often distilled into a single model for realtime inference.
VAD (Voice Activity Detection) VoiceLab: A classifier that returns 1 for speech and 0 for non-speech, frame by frame. The first stage of nearly every voice pipeline, from conferencing apps to podcast editors to the VoiceLab QA demo.
Voice cloning VoiceLab: Synthesising speech in the voice of a target speaker from a reference recording; modern models work from 3–30 seconds of reference audio. The basis for both legitimate dubbing/accessibility tools and modern impersonation attacks.
WDRC HearLab: Wide Dynamic Range Compression: the multiband, level-dependent gain processing inside every modern hearing aid. Maps the speaker’s dynamic range into the user’s residual hearing range.
WebAudio API MixLab: The W3C standard for audio processing in browsers. Includes AudioContext, AudioNode, AudioBuffer, AnalyserNode, AudioWorklet, and OfflineAudioContext. The foundation under every browser-side audio tool.