‘VoiceGate’: Accentize's Concept of Machine Learning Algorithm for Real-Time Noise Reduction

Based on machine learning techniques, ’VoiceGate’ smartly cleans up various noises from speeches and vocal recordings in real-time.

The Music Telegraph | Text 2020/03/25 [10:43]

‘VoiceGate’: Accentize's Concept of Machine Learning Algorithm for Real-Time Noise Reduction

Based on machine learning techniques, ’VoiceGate’ smartly cleans up various noises from speeches and vocal recordings in real-time.

The Music Telegraph| 입력 : 2020/03/25 [10:43]

 

▲ VoiceGate

© Accentize



Recent advances in the area of machine learning have evidently shaped the way we handle, understand, and process data. New theoretical insights, the rise of freely available programming libraries, and the increasing access to computational resources provide a huge set of new possibilities to tackle data processing problems from a different angle. Accentize focuses on applying cutting-edge machine learning methods to the area of audio signal processing and closes the gap between state-of-the-art research and productive implementations. The newly released Accentize’s ‘VoiceGate' plug-in will help you to clean up almost all possible noises from your speech and vocal recordings in an easy and automated fashion. 

 

 

‘VoiceGate’ is a real-time noise reducer working under specially trained artificial neural network that analyze human speech. The artificial neural network applied to ‘VoiceGate’ has been trained on more than 100 hours of audio data to learn the characteristics of human speech. With the help of machine learning techniques such as artificial neural networks, the algorithm of ‘VoiceGate’ can differentiate between desired signal components and unwanted noise which can be easily suppressed.

 

 

 

The VoiceGate Engine 

‘VoiceGate’ is designed to detect general noises in two types, which are ‘Steady Noise’ and ‘Impulsive Noise’. Under this concept of noise detection, steady noise means a stationary and constant noise which does not fluctuate over time in its amplitude level such as white or pink noise. On the other hand, impulsive noise stands for a transient and instant noise which unexpectedly occurs over time such as a click or pop noise. Once the noise has been categorized into one of these two types, then ‘VoiceGate’ starts its  attenuation processing against the noise over broad frequency range or only specific frequency region which users can define through the control panel of ‘VoiceGate’. 

 

 

 

Controls Over Noises

Don’t forget the ‘VoiceGate’ is running under machine learning algorithm, so users do not really need to indulge in controlling parameters because ‘VoiceGate’ almost automatically implements controls over the noises. However, we need to briefly explore the functions and controls of ‘VoiceGate' in order to get better results from the machine learning mechanism of controlling noises. 

 

 

▲ Mode selection: Broadband Mode, Spectral Focus Mode


At the top of the plugin, there are two mode-selection buttons, which are ‘Broadband Mode’ and ‘Spectral Focus Mode’. The default is set to Broadband Mode, and under this mode ‘VoiceGate’ lets users to control noises over the entire range of frequencies that ‘VoiceGate’ can cover (approximately 50 Hz to 10,000 Hz). 

 

 

 

▲ The upper window shows the current frequency response of the filter


In the center of the interface two windows are being displayed. The upper window shows the current frequency response of the filter covering over the whole frequency range which ‘VoiceGate’ can provide. Users can directly monitor the areas of frequencies being attenuated by the filter and the degree of the attenuation through this window. 

 

 

 

▲ The lower window shows how much overall signal passes through the plugin and how much gets filtered out


The lower window shows the output waveform along with the gray input waveform in the background. Here, it is possible to observe how much overall signal passes through the plugin and how much gets filtered out in real-time. The gray waveforms indicate the parts that has been cut through noise filtering.

 

 

 

▲ Parameters for Steady Noise and Impulsive Noise


At the bottom of the interface, there are two different control sections for two different noise types with two parameter controls in each. The left section is for controlling steady noises which do not change much overtime and the right section tackles impulsive noises such as click or pop noises. For both sections, it is possible to define a maximum allowed attenuation and a sensitivity. The ‘Max Reduction’ sets the maximum allowed gain reduction of the selected noise. The gain reduction will never fall below the maximum allowed attenuation in each section. If you want to only have steady or impulsive noises to be filtered out, you can set the maximum attenuation while deactivating the other section by setting the gain reduction to 0 dB. The ‘Sensitivity’ defines how aggressive the filter will tackle noises. A low percentage means very little noise-reduction effect. These parameters can be used to find a good trade-off between attenuated noise and arising artifacts. Usually the right Impulsive Noise section should work well with most noises. However, sometimes if there is a persistent and constant noise it can result in fast modulated noise due to the rapid gain changes. Here the Steady Noise section comes in handy to get more pleasing results.

 

 

 

▲ Spectral Focus Mode


At the top of the plugin, you will find the Spectral Focus selection along with the Bypass button. Spectral Focus Mode allows users to individually set parameters for a specific frequency region. You can select one zone by clicking from the three editable zones and set the range of the selected zone by dragging the white dot to the left or right, and then apply the parameter values to the selected zone. The default is set to all the three frequency regions to be processed at the same time, but, for example, if you want to attenuate only some low frequency noises without touching high frequencies which contain consonant sounds then you can select the low frequency region to control only those low frequency noises. In addition, You can adjust the cross-over frequencies of the different bands by dragging the white dots as you wish in the upper window of the interface. In the parameter section, there are six different knobs for each noise type. These controls work the same way as in the Broadband Mode but now for each band individually. If you want no reduction in one band at all then set the Max Reduction parameter to 0 dB. 

 

 

 

The Sound from 'VoiceGate'

Under Broadband Mode, ‘VoiceGate’ reacted instantly against almost all types of noises from the vocal track recorded at a poorly treated room in acoustically. The Steady Noise filter automatically detected and removed intermittent birds singings coming through the window and the unknown hum noises. Meanwhile, the Impulsive Noise filter eliminated any sibilant and harsh breathing sounds from the vocal recorded with a low-cost microphone in real time. ‘VoiceGate’ removes noise in this way reproducing sounds more clear and definite, and the surprising thing is that the sound processed by VoiceGate has no distortion or exaggeration due to keeping the frequency areas of sounds other than noise intact. This means the filters do not affect the areas outside the noises.  

 

Spectral Focus Mode enables surgical elimination of noises in a narrow frequency area, so I could remove only high-frequency sibilances from the vocal keeping the other noises still alive. Users can monitor noises from the selected region and this seems quite effective in finding target-noises to be removed or preserved. 

 

The VoiceGate’s machine learning algorithm will be continuously improved by constantly adding new data from various user experiences into its artificial neural network, so ‘VoiceGate’ will handle almost all kind noises from human speeches and vocals.

 

 

 

 

 

With ‘VoiceGate’ You can do:

 

- Reduce noises from speech and vocal recordings in real-time

- Eliminate clicks and pops

- Eliminate background noises

- Tackle different kinds of noises without introducing many artifacts

- and more.

 

 

 

 

Price:

 

€84 EUR

 

 

*Note that a 7-day, fully-functional trial version for macOS and Windows can be downloaded free: here

 

 

 

 

Accentize Recent advances in the area of machine learning have evidently shaped the way we handle, understand and process data. New theoretical insights, the rise of freely available programming libraries and the increasing access to computational ressources provide a huge set of new possibilities to tackle data processing problems from a different angle. We focus on applying cutting-edge machine learning methods to the area of audio signal processing and close the gap between state-of-the-art research and productive implementations.

 

 

 

For more information on 'VoiceGate'

 

 

View this article in Korean version:  1   2

 

View this article in Japanese version:  1  

 

 

Comment
  • 도배방지 이미지


광고
광고
광고
광고
광고
광고
광고
광고
광고
광고
광고
광고