Audio

File info, waveform and channel readout#

What it does. Decodes the audio (Web Audio decodeAudioData) and shows file info, container/codec details, per-channel peak/RMS, and a waveform.

How to reach it. Drop any recognised audio file. Built in web/assets/js/renderers/audio.js.

How to use it. For multi-channel audio (audio.js's channel selector), choose Mix (every channel averaged) or an individual channel (Left, Right, Centre, LFE, surrounds - named per the file's declared speaker layout, best-effort) to drive the spectrogram and waveform; per-channel peak and RMS update with the choice. describeChannels()/channelLabels() recognise mono through 9.1.6 Atmos layouts.

Spectrogram#

What it does. An interactive time/frequency/loudness plot of the audio, computed with a hand-written FFT (spectrogram.js's fft(), no audio library dependency).

How to reach it. Automatic below the waveform for any decoded audio.

How to use it. Linear or logarithmic frequency axis, adjustable FFT size and window function, choice of colourmap, and Save PNG to export the canvas. computeReassignedSpectrogram() offers a sharper reassigned-spectrogram mode; computeStftComplex()/ combineStftToDb() support blending two signals' spectra (used by the AI separation blend slider below).

The frequency axis is a segmented toggle (the selected option carries the accent fill):

Notes / limits. Reaches down into the sub-20Hz range per the repo root README.md.

Codec / loudness / pitch / tempo analysis#

What it does. Peak/RMS level stats, spectral centroid, LUFS integrated loudness, pitch detection (YIN algorithm), BPM/tempo detection, and stereo correlation metrics - all pure computation in audio-analysis.js over the decoded sample buffer.

How to reach it. Automatic - shown in the audio info readout.

How to use it. BPM is read from ID3/MP4 tags when available (audio-codec.js's readTagBPM()), otherwise estimated via onset detection (detectBPM() in audio-analysis.js).

Notes / limits. No interactive controls - a readout only.

Signal health and musical key (folded into File info)#

What it does. Adds four pure-computation reads to the File info card: crest factor (peak-to-RMS, a dynamics/loudness-war gauge), DC offset (non-zero mean = capture/hardware fault), effective bit depth (the deepest bit actually carrying signal, recovered from LSB activity - exposes padded/upscaled "hi-res" files), and musical key (chroma matched against the Krumhansl-Schmuckler key profiles, with confidence and a runner-up).

How to reach it. Automatic, in the File info readout. Built in audio-forensics.js (signalHealth, detectKey).

Notes / limits. Key detection is most reliable on tonal music; the runner-up is often the relative major/minor. Effective bit depth compares against the declared depth and flags a likely padded/upscaled file.

Advanced (forensic panels)#

What it does. A single collapsed Advanced card - mirroring the photo and video Advanced cards - that gathers the deep forensic reads into collapsible <details> panels, keeping them out of the everyday File info readout above. Each panel below is closed until opened:

• Loudness meter (EBU R128) - the full broadcast/streaming loudness set beyond the single integrated-LUFS row: gated integrated loudness (ITU-R BS.1770), momentary (400 ms) and short-term (3 s) maxima, Loudness Range (LRA), and true peak (4x-oversampled dBTP, catching inter-sample overs that sample-peak metering misses), plus a loudness-over-time plot with a -14 LUFS streaming-target reference line. The plot is playable like the waveform - click or drag to seek, own transport below - and its playhead is measured against the decoded buffer's duration, the same time base the curve is plotted in and the one every other playhead on the page uses, so all of them stay in step. Measured on the channel-merged signal; a true peak above 0 dBTP is flagged in the accent colour (delivery specs cap it at -1 dBTP).
• Lossy-source check - decides whether a file that claims to be lossless (FLAC/WAV/ALAC...) was really made from an MP3/AAC, by finding the hard spectral low-pass a lossy codec leaves and mapping its cutoff to a probable source bitrate. A genuine lossless file reaches ~95%+ of the Nyquist limit.
• Mains hum / ENF - narrowband 50 or 60 Hz energy (plus harmonics) from power-line interference; reports the exact frequency and implied region - the gateway to ENF forensic timestamping.
• Ultrasonic content - energy and any narrowband tones above ~18 kHz (tracking beacons, device-pairing tones, watermarks), when the sample rate is high enough to carry them.
• Touch-tones (DTMF) - phone digits decoded via Goertzel filters at the 8 standard row/column frequencies, with per-digit timing (shown only when tones are found).

How to reach it. Automatic collapsed card below File info for any decoded audio (the card is omitted only if no panel has anything to show). The lossless spectral panels derive from one long-average (Welch) spectrum of the whole file. Built in audio-forensics.js (loudnessR128, truePeakDb, analyzeTranscode, analyzeMainsHum, analyzeUltrasonic, detectDtmf); rendered by audio.js (the aAdvPanel helper mirrors video's vAdvPanel).

Notes / limits. The lossy-source cutoff and its bitrate mapping are approximate (encoders and settings vary); a "likely lossy source" verdict on a declared-lossless file is a strong tell, not proof. Nothing is uploaded - all analysis is on-device.

Playback transport#

What it does. A custom <audio> transport shared across the audio, video, and sonify players: play/pause/replay button, a draggable seek track, and a current/total time readout.

How to reach it. Automatic wherever audio plays. Built in audio-player.js's makePlayer(), re-exported by audio.js and reused by video.js.

How to use it. Click the transport button to play/pause, or (once ended) replay from the start; drag the seek track to scrub. The volume button opens a popup with a vertical slider (0-225%, so audio can be boosted past its native level) and mute toggle - volume is shared across every player on the page via a small registry (registerVolPlayer), so setting it once applies everywhere. The slider is logarithmic: hearing is roughly logarithmic too, so position maps to gain through a dB scale (-40 dB up to the 100% tick, then 0 to +7 dB above it) rather than straight through. A linear track spends most of its length on the top few dB and leaves no room to set a quiet level.

Frequency isolation (band-stop EQ)#

What it does. An interactive parametric EQ that solos or cuts specific frequency bands, with one-tap presets and free-form custom bands.

How to reach it. Click Isolate in the audio actions row (only offered when driving file playback, not live capture). Built in audio.js.

How to use it. Click + Custom band to add a band, drag/edit its range; each band gets its own remove (x) button. Preset buttons apply a named frequency preset as a one-tap solo (cutting everything outside the preset's range); Clear removes the active preset. Once at least one band exists, Download WAV renders and downloads the isolated result.

AI vocal separation#

What it does. Splits a track into separate vocal and instrumental stems using an on-device neural network (MDX-Net "Kim Vocal 2" from Ultimate Vocal Remover), entirely in-browser via ONNX Runtime Web.

How to reach it. Click Separation in the audio actions row under the spectrogram (next to Isolate) to open the AI options row, then click Standard, Lite, or Denoise. Lite is selected by default on phones and lower-memory devices. Built in audio.js, backed by the MDX-Net subsystem in web/assets/js/lib/mdx-*.js (see parsers-and-libs.md).

How to use it. Click Standard or Lite (mobile-friendlier) to separate with that model. First use prompts to confirm the model download (Download and continue/Start separation to proceed, Cancel to back out) unless already cached. Once separated, each stem gets Play (in a blend row that can mix vocal/instrumental live), Analyse (runs the stem through the full audio analyser), and Download WAV. A [?] info button next to "Separate" explains the approach. The model options collapse and Separation is deselected when the result is ready. Press Separation again to choose a model and confirm a new run against the original file; cancelling keeps the current result.

Pick a model tier, then confirm the one-time model download:

Each separated stem gets its own controls:

Notes / limits. Runs on GPU via WebGPU where available, WASM otherwise. WebGPU automatically falls back to WASM if initialisation or inference fails; a stalled/crashed GPU worker is also replaced and retried once on WASM. WebKit loads the smaller WASM-only runtime, reducing fixed runtime memory on older iPhones while retaining ORT's normal inference allocation strategy. Download, cache, runtime initialisation, audio preparation and inference are reported as separate progress phases. The first inference window has a distinct starting state because no measured percentage exists until it returns; later windows show a matching percentage and bar. A completed download is therefore not mistaken for a completed separation. Model downloads and cached entries are checked against the pinned byte size, and a bad cached object is removed before retrying. The Standard model and both runtime variants (~96MB total) are part of the "Complete" offline tier (see pwa-offline.md). The isolate band-stop cuts are re-applied to separated stems in parallel (nodes can't be shared across the file player's audio context and the stem-blend context), so Isolate edits and Separation compose rather than one bypassing the other. Whole-song padding, per-window FFT allocations and eager WAV copies are avoided; the model is prepared before song-sized channel copies are made, and WebKit releases the AI worker after returning a result. A stem WAV is encoded only when its Play or Download button is used, which reduces peak memory on phones.

AI denoise#

What it does. Removes background noise and hiss from a track while keeping the full sound, using an on-device neural network (DeepFilterNet3, a 48 kHz full-band speech/audio denoiser), entirely in-browser via ONNX Runtime Web. It produces two stems - Clean (the enhanced audio) and Noise (what was removed, = original - clean) - shown in the same blend view as vocal separation, relabelled Clean to Noise.

How to reach it. Open the AI options row by clicking Separation in the audio actions row under the spectrogram (the row is closed by default), then click denoise to the right of the Standard/Lite tiers. Built in audio.js, backed by the DeepFilterNet subsystem in web/assets/js/lib/dfn-*.js (see parsers-and-libs.md).

How to use it. First use prompts to confirm the one-time model download (Download and continue/Start denoise to proceed, Cancel to back out) unless already cached. Once done, the blend slider fades Clean to Noise, and each stem gets Play, Analyse and Download WAV, exactly like the separation stems. Denoise and vocal separation are mutually exclusive - starting one clears the other, since only one blend owns the spectrogram at a time.

Notes / limits. Unlike separation, this is noise suppression, not source separation - it is aimed at cleaning up recordings (voice, ambience, hiss), not splitting a musical mix. It deliberately uses single-threaded WASM on every browser because the recurrent graph produced incorrect output on the WebGPU backend. The immutable, exact-size-validated model (~9MB on top of the shared ONNX runtime) is part of the "Complete" offline tier (see pwa-offline.md). All the DSP around the model (STFT, ERB features, the gain mask and the deep filter) is reimplemented on-device in dfn-dsp.js/dfn-enhance.js, mirroring libDF; long audio is processed in overlapping segments to bound memory.

Waveform selection: zoom and export#

What it does. Select a time range on the waveform to zoom into it or export just that range.

How to reach it. Drag a selection on the waveform, then click Zoom or Export WAV. Built in audio.js. Reset zoom returns to the full view.

Reversed playback#

What it does. Plays and downloads the decoded audio backwards - each channel's samples flipped, re-encoded as WAV.

How to reach it. Click Reverse audio ("Reverse" card, ↺ icon) below the audio analysis. Built in media-reverse.js's buildReverseAudioCard().

How to use it. Click ↺ Reverse audio; once rendered, a player appears with Download reversed (WAV) and Analyse reversed (feeds the reversed clip back through the full analysis pipeline via window._anrHandleFile).

Notes / limits. The reverse itself is synchronous and near-instant but gated behind a click so it isn't computed for every file automatically. Reuses the AVI module's PCM-WAV encoder (encodeWav, from video-avi.js). Video reverse is a separate feature (FFmpeg re-encode) - see video.md.

Microphone recording#

What it does. Records live audio from the microphone and runs it through the same analysis as a dropped file.

How to reach it. Click Record on the dropzone. Built in audio.js.

How to use it. Shows a compact live streaming spectrogram while recording (a lighter-weight visual than the full interactive one). Click Stop to end the take; Download recording saves the captured audio.

Live spectrogram (no recording)#

What it does. Visualises the microphone input live, without recording - a real-time spectrogram of whatever the mic hears.

How to reach it. Click Live spectrogram. Built in audio.js.

How to use it. Much the same toolbar as the file spectrogram (Save PNG, Fullscreen), plus its own Record and Live spectrogram toggle - which the file spectrogram does not carry - and Pause and Analyse last Ns, which freezes and analyses the trailing N seconds of the live stream. A channel picker (Mix/L/R/etc.) chooses which input channel feeds the view.

Tracker modules (MOD, XM, IT, S3M and ~60 more)#

What it does. A tracker module is a score, not a recording: a bank of instrument samples plus a pattern grid saying which sample to trigger, on which channel, at which row. Analyser renders the song to audio on your device with libopenmpt - the engine behind OpenMPT - and then treats the result as ordinary audio, so a .mod gets the entire Sound section: waveform, spectrogram, transport, loudness, key and BPM. Above it sits a tracker card with the tracker that wrote the file, the channel, pattern and order counts, the song message, and the sample and instrument name lists, which is traditionally where authors left their greetings. The rendered song downloads as a WAV.

How to reach it. Drop any of the ~60 supported module extensions. Built in tracker.js on lib/openmpt-loader.js.

Notes / limits. libopenmpt is ~1.5 MB and lazy-loaded, in the Everything offline tier. Rendering is bounded by output length rather than file size - a 40 KB .it can be a twelve-minute song, and a module written to loop has no end at all - so a very long or looping module is rendered up to the ceiling and the viewer says so. .mod also names a JVC camcorder video, which is routed away on its bytes.

Cover art / tag extraction#

What it does. Reads ID3/MP4 tags, including embedded cover art, from the audio file.

How to reach it. Automatic - built in audio-codec.js's extractCoverArt()/readAudioTags().

Notes / limits. The tags are a readout. Embedded cover art also gets an Analyse cover art box at the end of the analysis, which opens the picture in the photo tools.

ADTS-to-M4A wrapping#

What it does. Wraps raw AAC (ADTS) streams in a minimal M4A container so browsers that won't decode bare ADTS can still play the file.

How to reach it. Automatic when the file's container/codec is sniffed as raw ADTS (audio-codec.js's peekContainer()/adtsToM4a()).

Notes / limits. Internal compatibility shim; no user-facing control.