Need an algorithm (or a set of good ones to compare for various input data) that will reduce the noise level of a voice audio signal without distorting the signal appreciably using Java.
Input is an audio signal that includes a voice along with some background noises. The noise varies over the course of the recording. There definitely exists ways to remove noise like this, developed for voice recognition and movie making.
The desired output is a minimally distorted voice signal with the background distractions minimally audible to the human ear. The quantitative criteria are minimized
- Signal to noise ratio and
- Total harmonic distortion.
You are looking for adaptive noise removal and possibly a variety that adapts to changing noise conditions over time within the same stream or file.
Older approaches include:
- Remove frequencies containing the majority of noise using bandpass and/or notch filters (which only works well if the desired signal and the noise band do not intersect)
- Drop the noise level at points between words, notes, or other audio events (in the dead space) as does the Dolby noise reduction system scheme
- Drop the noise floor across an entire file using a Hamming window or other window in conjunction with an FFT library
- Hand edit sections of an audio track in programs like Cakewalk or its competitors
These methods have been found less than desirable when trying to clean up a larger file or multiple files or in real time applications such as real time voice recognition or telephony.
One of the Java programs that I have not personally tried his here. Even though it has some level of automation, it is an LSE (least squares estimator) which works across a block of data but is not suitable for continuous operation or an audio file with changing unwanted noise conditions. (It's not as adaptive as one might hope.)
The solution that I found after much investigation and now use all the time is not written in Java. It is a MATLAB program that can also run in open source Octave with minimal modification. I started porting it to C++ but ran out of time to finish that.
The class of algorithm it implements is called MMSE (Noise Reduction Using Minimum Mean Square Estimators). The MATLAB version has been refined several times by Dr. Hendricks, concluding with the the 2010 version.
I've compared it with its competitors for both dialog and music and it equals or excels beyond the others in all cases I've tried. (I have no affiliation with Dr. Hendricks or MATLAB other than I like the results I've been getting from his implementation on that platform.)
来源：https://stackoverflow.com/questions/42314122/audio-noise-removal-in-java