Pseudo real-time content-aware auditory cleansing

1 . A method for pseudo real-time content-aware auditory cleansing, the method comprising:

introducing a look-ahead delay in delivering an audio file by an audio delivery system, the look-ahead delay being a length of time that is largely imperceptible to a listener;

collecting, by a buffer in a sidechain, a sample of the audio file, the sample having a length equal to the look-ahead delay;

predicting, by machine learning-based phoneme classifier, a presence of a sibilance in the sample in real time, the machine learning-based phoneme classifier receiving the sample from the buffer and generating a classified sibilance type within a timeframe that is equal to or less than the look-ahead delay, the machine learning-based phoneme classifier being able to predict which type of sibilance from a sibilance classification set that includes {s, sh, t} is present and to output a prediction selected from the sibilance classification set as the classified sibilance type;

cleansing, in the sidechain, the sample of the presence of the classified sibilance type; and

outputting a cleansed version of the sample.

2 . The method of claim 1 , wherein the cleansing the sample comprises compressing the sample according to a gain control setting corresponding to the classified sibilance type, the gain control setting determined by a gain computer.

3 . The method of claim 2 , wherein the gain control setting comprises a gain reduction.

4 . The method of claim 1 , wherein the look-ahead delay and the length of the sample each are less than 15 seconds.

5 . A method for pseudo real-time content-aware auditory cleansing, the method comprising:

collecting, by a buffer in a sidechain, a sample of the audio file, the sample having a length equal to a look-ahead delay, the look-ahead delay being a length of time that is largely imperceptible to a listener;

predicting, by a machine learning-based classifier, a presence of an unwanted sound in the sample in real time, the machine learning-based classifier receiving the sample from the buffer and generating a classified unwanted sound type within a timeframe that is equal to or less than the look-ahead delay, the machine learning-based classifier being able to predict which type of unwanted sound from a classification set that includes {s, sh, t} is present and to output a prediction selected from the classification set as the classified unwanted sound type;

cleansing, by a functional module in the sidechain, the sample of the presence of the unwanted sound; and

outputting a cleansed version of the sample.

6 . The method of claim 5 , wherein the cleansing the sample comprises compressing the sample according to a gain control setting corresponding to the classified unwanted sound type, the gain control setting determined by a gain computer.

7 . The method of claim 6 , wherein the gain control setting comprises a gain reduction.

8 . The method of claim 5 , wherein the cleansing the sample comprises an equalization adjustment according to a control setting corresponding to the classified unwanted sound type.

9 . The method of claim 5 , wherein the cleansing the sample comprises removal of the presence of the unwanted sound.

10 . The method of claim 5 , wherein the cleansing the sample comprises replacement of the presence of the unwanted sound.

11 . The method of claim 5 , wherein the unwanted sound comprises one, or a combination, of a sibilance sound, a plosive, a breath, and a non-speech sound artifact, the machine learning-based classifier able to determine whether the unwanted sound comprises the sibilance sound, the plosive, the breath, and the non-speech sound artifact.

12 . The method of claim 5 , wherein the look-ahead delay and the length of the sample each are less than 15 seconds.

13 . The method of claim 5 , wherein the functional module comprises one, or a combination, of a gain computer, an equalizer module, and a sound replacement module.

14 . A system for pseudo real-time content-aware auditory cleansing, the system comprising:

a buffer configured to collect a sample of an audio file, the sample having a length equal to a delay being introduced in the delivery of the audio file, the delay comprising a length of time that is largely imperceptible to a listener;

a machine learning-based sound classifier comprising configured to receive the sample from the buffer in a sidechain and output a prediction of a presence of a classified unwanted sound type in the sample within a timeframe that is equal to or less than the look-ahead delay, the machine learning-based classifier able to determine whether the unwanted sound comprises one or more of a sibilance sound, a plosive, a breath, and a non-speech sound artifact, the machine learning-based sound classifier being able to predict which type of unwanted sound from a classification set that includes {s, sh, t} is present and to output a prediction selected from the classification set as the classified unwanted sound type; and

a functional module configured to cleanse the sample of the classified unwanted sound type, the functional module further configured to output a cleansed version of the sample.

15 . The system of claim 14 , wherein the functional module comprises one, or a combination, of a gain computer, a compressor, an equalizer module, and a sound replacement module.

16 . The system of claim 14 , wherein the audio file comprises a vocal track.

17 . The method of claim 14 , wherein the functional module is configured to perform a control function configured to cause a cleansing action.

18 . The method of claim 17 , wherein the cleansing action comprises one, or a combination, of a compression, an equalization adjustment, a replacement, and a removal of the unwanted sound.