A simple frequency based Voice Activity Detection (VAD) algorithm written in C++.
This class is designed to replace with silence noise section of a recorded speech. The recorded speech must follow specific guidelines:
- It must be PCM encoded with 8-bit signed values, sampled at 8000Hz with 160 samples every 20ms.
- It must be saved in ./inputData/inputaudio
<num>.data where<num>is a placeholder for the actual int value
The output will be saved with the same encoding in ./outputData/outputaudio<num>.data where <num> is a placeholder for the value
previously submitted.
The actual functioning is rather simple:
- An input and output streams are opened in order to simulate a real time connection
- Every packet composed of 160 samples (minus the last one which is allowed to have less samples) is processed as follow: 2.1 If in the last two packets there was a voice-match it will be transimitted regardeless of its content 2.2 Using FFT (Fast Fourier Transformation) a packet is categorized as either voice-match or not thanks to its main frequency 2.3 If it is a voice-match then: 2.3.1 The most recent packet was also a voice-match then just the current packet is sent 2.3.2 The most recent packet was not a voice-match then both packets will be sent 2.4 If it is not a voice-match then: 2.4.1 If the most recent packet was a voice-match then a silenced packet (composed of zeros) is sent
- Both input and ouput streams are closed to prevent undefined behaviours
In fact, quite well. It is suggested to download a music player such as Audacity and use the "import raw data" option in order to specify sampling rate (8000Hz) and encoding (PCM - 8 bit signed). All audio files are already available in the "outputData" folder.
Compile the source code with g++ *.cpp *.hpp -o vad in order to get an executable.
The next step is just to use it with ./vad and follow the instructions.