@audio/denoise vs Neiro

@audio/denoise is a free web tool for cleaning recordings and listening to or saving the result. It handles targets including steady background noise, mains hum, clicks, pops, vinyl crackle, clipping, wind rumble, plosives, sibilance, room reverb, and noise between phrases. The page describes thirteen open-source restoration methods and a classifier. Automatic mode measures properties of a recording, selects a specialized method, and displays its plan; users can instead choose methods such as dehum, declick, declip, and gate. Audio is processed in the browser, and the page says it is not uploaded or sent off the device. The approach uses classical signal processing rather than AI noise removal. For music, the page identifies hum, clicks, crackle, and clipping as treatable, while advising light use of broadband denoising and comparison with the original. Machine-learning models may perform better on heavy speech-over-noise. The associated npm package processes mono Float32Array audio. Automatic mode does not choose dereverb, which must be selected explicitly.

Neiro is a local audio worksuite for separating sources, restoring recordings, transcribing audio and editing waveforms. Processing takes place on the user's machine and audio does not leave it. Neiro offers a Tauri desktop app and a browser interface launched with `neiro ui`; desktop installers are listed for Windows, macOS and Linux. Separation covers vocals, instrumentals, harmonic and percussive parts, four- or six-stem mixes and drum kits, with a null-test residual for each result. Restoration tools include declipping, hum removal, denoising, dereverberation, bandwidth extension and reference mastering. Transcription can export MIDI, MusicXML, ASCII tablature and LRC lyrics. Studio provides non-destructive waveform edits, while Learn includes loop regions, count-in, metronome, step mode, WebMIDI and DAW wait mode. VST2 and CLAP injectors can capture audio from a DAW into Neiro. Core DSP works without model downloads, but optional neural backends download weights on first use. The Python package requires Python 3.10–3.12; compressed or video inputs also require ffmpeg on PATH.