Algorithms

Main UI

Preset configurations

Algorithm Description
DSP – I A noise probability distribution model is adapted based on the similarity of current short-term statistics to the initial noise model in a logarithmic spectrum domain and used to suppress the audio noise in the spectrum domain [1].
DSP – II Similar to DSP – I, but the noise estimation is done in mel-spectrum (auditory spectrum) domain.
VB – I NEW Vocal Boost; boost vocals and similar sounds while reducing noise.
SE – I A deep neural network speech enhancement model, trained to separate speech from background noise.
SE – II NEW A new speech enhancement method combining deep learning and DSP to efficiently separate speech from competing sounds while improving clarity and intelligibility.

SE-II Example

Original

SE-II processed

Spectrogram Editor

General Usage of the Editor

Some algorithms are available only on the Spectrogram Editor because they work best with precise time-frequency region selection.

Select the processing method and the intended input source (Original or Draft). Draw a selection to indicate the time-frequency region to be processed and adjust the method parameters as needed, then press the "Process" button. This creates a new Preview signal which can be inspected by listening. It can be accepted as the new current Draft by pressing "Keep changes", which also updates the Draft spectrogram.

Note: In order to process the uploaded audio from the beginning using the editor, rather than working on the latest output, "Original" must be selected as the input/source before the first processing phase.

Usage example

Spectral Erase

Steep FIR filtering suppresses unwanted stopband frequencies, while statistically matched replacement noise is generated using mean and variance estimates from the preceding short-time region. Suitable for removing transient noise events precisely in time and frequency. This algorithm processes each channel of the input.

Spectral Floor

An elevated version of the spectral floor profile is used as a reference for attenuating frequency bins that exceed it, helping to smooth out transient events while maintaining the stable background.

Spectral Subtraction

Continuously tracked minimum statistics are used to estimate the background noise floor and its mean spectral shape. Spectral subtraction is then applied to reduce the slowly changing background noise.

Vocal Boost

Following initial noise reduction, acoustic resonances are further amplified using a suitable IIR filter, making features such as speech formants more prominent.

This algorithm is also available on the Main UI, but the Editor version also allows tuning the level of vocal boost ("boost amount").

References

  1. Log-spectral and cepstral noise reduction using an adapted noise model

Main page     Top

Thu Nov 12 17:04:17 UTC 2026