Skip to content

KudioStudio

KudioStudio

KudioStudio is a desktop workbench for kudio, my Python toolkit for speech audio. Open a clip, look at it, change it, measure it and build a dataset from it, without writing a script first. The panels run documented kudio functions, and each one shows the call its controls make. That line is written to be read; what an edit records is the exact call, with the seed a random effect actually drew and the full path of any file it reads, so whatever you settle on here carries straight over to code.

The KudioStudio window

Install

pip install kudiostudio
kudiostudio

It needs Python 3.9 or newer. Getting started lists the extras and what Linux needs, and walks through a first session from opening a file to saving it.

The panels

The rail on the right of the window holds one tool per panel:

Panel What it is for kudio behind it
Inspect the analysis view's settings, the selection, and the Levels readout STFT, melspectrogram, mfcc, audio_report, loudness_over_time, true_peak
Effects one effect at a time on the selection or the whole clip, heard before it is applied kudio.effects, rir, apply_rir, rt60
Enhance seven statistical denoisers over four noise estimators, trad_enhance, a wavelet filter, and learned models spectral_enhance, trad_enhance, wavelet_low_pass_filter, kudioenhance
Compare the seven statistical denoisers, the four noise estimators or the whole grid on the same clip, ranked, with a learned model as one more row if you give it one; each row is a click from your ears compare_enhancers
Record a fixed-length take, or one that runs until you stop it, drawn as it arrives record, StreamRecorder, SpectrogramStream, StreamEnhancer
Synthesize a noisy dataset from a clean folder and a noise folder, rooms and phone lines included Synthesizer
Manifest that dataset read back: missing files, and the noise × SNR coverage load_manifest
Pitch the f0 contour over the spectrogram, and the voiced spans as markers f0
Metrics what is wrong with a recording when there is no reference, and scores when there is audio_report, dnsmos, find_delay, eval_metrics, snr, si_sdr, segmental_snr
Batch a whole folder converted, normalised, denoised or given a reverberant copy, with a report of the files that missed convert_folder, enhance_folder, rir, apply_rir

Where to go next

  • Getting started

    Installing, the extras, the command line, and a first session from opening a file to saving it.

  • The window

    What each part of the window is, how to select, zoom and play, and every shortcut.

  • Editing and markers

    Cut and paste with a crossfade at every join, A/B against the original, markers that outlive the session, and your edits written out as a script.

  • Panels

    One page per tool on the rail, with every control and what it calls.

  • Writing a panel

    A tool of your own on the rail, from a package of its own.

  • Troubleshooting

    What it means when the window, the sound or a panel does not do what you expected, and what to do about it.

kudio
The toolkit underneath: files, features, effects, rooms and channels, dataset synthesis, denoising, metrics and recording. Its own guide is at recklight.github.io/Kudio.
kudioenhance
Trains the learned denoisers that the Enhance and Compare panels load.