Auditius

Auditius is an audio editor and multitrack studio that runs in the browser, in the spirit of Cool Edit Pro, the editor I used in 2002 until it stopped working and that Adobe bought in 2003 to turn into Audition. Plain menus that do what they say, a waveform you edit directly and a multitrack view next to it. It is free and there is no account.

The audio never leaves the computer

That is the rule everything else follows. Opening a file, applying an effect, reducing noise, stretching the tempo and encoding the result all happen on the device. There is no upload step because there is no server to upload to: the site is a set of static files.

The work is saved in the browser as you go. Close the tab and come back the next day, and the files, the cues and the multitrack session are where you left them. If the editor is open in two tabs, only one of them saves and the other says so, with a button to carry on there instead.

Editing a waveform

The edit view is the classic destructive editor:

• Waveform and spectral display, with selections that can take a single stereo channel, the way Cool Edit did it.
• Cut, copy, paste, mix paste and trim, snapped to zero crossings, with undo as deep as memory allows.
• 38 effects with presets and live preview: EQ, compression, reverb, delay, chorus, flanger, phaser, distortion, pitch and tempo.
• Restoration: noise reduction from a captured profile, click and pop removal, hiss reduction and clip restoration.
• Cues, scrubbing, a pencil to redraw samples, recording from the microphone, and scripts that replay a chain of effects over many files.

The multitrack

Blocks loop by dragging their right edge, fade with handles and crossfade where they overlap. Each block has volume and pan envelopes, each track has a real-time effects rack and automation lanes, and buses with sends let several tracks share one reverb. A built-in library of 356 loops in twelve genres, from drums and bass to synths and effects, follows the session tempo, and a tempo map moves every block that follows it when the tempo changes. The mix goes out as WAV, FLAC, MP3 or Ogg, with the reverb tails allowed to ring out.

Songs from a seed

Generate ▸ Song composes a complete track for the multitrack in five styles: dance, trance, house, techno and reggaeton. It writes the chord progression, a melody built from a motif that repeats, answers and resolves, a bass line and an arrangement with builds, breaks and drops. Every part is played by instruments synthesized in code, with no samples and no model, and the drums and effects come from the loop library.

The panel offers the key, the tempo, the form and an instrument for the melody, the chords and the bass, and anything left on Auto is chosen by the seed. The generator is deterministic: the same choices with the same seed always give the same song, title included. File ▸ Open Demo starts with a song arranged by hand and a few generated ones.

That is what makes songs shareable without a server. A generated song travels as its options and seed, a link of about 80 characters, and the browser that opens it composes it again. A song edited in the multitrack travels as its arrangement, compressed into the link. Recordings and imported audio cannot travel that way, and the dialog says which files stay behind.

AI that runs on your machine

The AI is optional and local too. Every model is downloaded only when you click Download, kept in the browser cache and loaded from there, with any other network request blocked while it runs.

• Transcription with Whisper, with a time for every word: clicking a word plays from it, and the text becomes SRT or VTT subtitles or one cue per sentence.
• Text to speech with Kokoro, Piper and Chatterbox, in Spanish and English, with pause tags in the text.
• Voice cloning with Chatterbox, gated by a consent checkbox. Saved voices stay in that browser, and generated audio is marked as synthetic in the file's metadata.
• Stem separation with Demucs, which splits a song into vocals, drums, bass and the rest, each as a new file, or into the vocals and the backing track for karaoke.

How it is built

Svelte 5 and TypeScript on the Web Audio API, with no backend at all. The landing pages are rendered at build time, so search engines read plain HTML, and the editor is only fetched when someone opens it.

Heavy work runs off the main thread. Effects, time stretching and the encoders run in a Web Worker with progress and cancel, so the interface never freezes on a long file. MP3 and Ogg go through LAME and Vorbis compiled to WebAssembly; the FLAC encoder and the ZIP writer behind the session files are my own.

Undo stores only what an edit changed. Each step keeps the affected range of samples rather than a copy of the file, under a shared memory budget, which is what lets an hour-long recording be edited in a browser tab.

The models fit the browser's rules. They run on ONNX Runtime, on the GPU through WebGPU when there is one and on WebAssembly otherwise. Even the runtime's own 26 MB binary is downloaded like a model instead of shipping with the site, which keeps every deployed file under the 25 MiB limit of static hosts. Two fixes were needed on the way: Whisper's word timings arrived one token late from the library, and the Spanish voices needed phonemes, which come from a rule-based converter of my own that matches eSpeak NG on 99.9% of words.

It was built with Claude, as a test of fast AI-assisted development that keeps the codebase coherent. The layers can only import downwards and a test enforces it, and the project ships with 769 unit tests and 173 end-to-end tests.

Status

Online since October 2026, in English and Spanish.