can MFCC coefficients alone be used to do voice detection?

Question

Aim of my project is to differentiate between cases of silence, single active speaker and multiple ( simultaneous ) active speakers using as basic ( but effective ) logic as possible. Is it possible to achieve this without use of extensive machine learning algorithms ??

---

Actually, i don't need to do speaker "identification". An audio stream will be taken as an input and I have to just say whether that audio consists of a single speaker, multiple speakers or silence... I have to implement this in C (real time), hence machine learning (unsupervised) would increase the computational complexity.

Answer 1

In essence, yes.

You would have to use something like Dynamic time warping to measure the similarities between two given signals, as long as they vary in time or speed.

From what I can remember, you can use DTW for speaking identification. BUT it is important to note that for single word recognition and single speaker identification this method would work but would not be suitable for commercial or large scale applications.

The DTW algorithm calculates the distances from the two signals, based on the Euclidean distances.. Whilst studying for my undergraduate degree, I did a simple word recognition system (In C++) and the machine learning algorithms (for instance, Hidden Markov Models) provided a much higher success rate than DTW.

Feel free to take a look at some of the code that I wrote for this project. Here

double Euclidean_distance(const vector<double> &actual, const vector<double> &training) {

    double distance = 0.0;

    for(unsigned i=0; (i < actual.size()); i++)
    {
        distance = pow((actual[i] - training[i]), 2);
    }

    return sqrt(distance);
}

double Distance(const std::vector< std::vector<double> > &actual, const std::vector< std::vector<double> > &training)
{
    int m = actual.size();
    int n = training.size();

    double cost[m][n];

    cost[0][0] = Euclidean_distance(actual[0], training[0]);

    for(int i = 1; i < m; i++)
            cost[i][0] = cost[i-1][0] + Euclidean_distance(actual[i], training[0]);
    for(int j = 1; j < n; j++)
            cost[0][j] = cost[0][j-1] + Euclidean::Euclidean_distance(actual[0], training[j]);

    for(int i = 1; i < m; i++)
            for(int j = 1; j < n; j++)
                cost[i][j] = std::min(cost[i-1][j], std::min(cost[i][j-1], cost[i-1][j-1])) 
                    + Euclidean_distance(actual[i],training[j]);

    return cost[m-1][n-1];

}

(Note this is Academic code, not tested and just illustrate the algorithm)

Answer 2

From what I understand you want to do a Voice Activity Detector (VAD) (in a first stage). For this there are several implementations with distinct complexity. Here you can find some algorithms:

https://stackoverflow.com/questions/5367214/any-simple-vad-implementation

To estimate how many speakers are speaking, things get a bit more difficult. Maybe you could have spectral signatures for each speaker and try to find them. Or you can use fundamental frequency to detect one or more that one frequencies. This problem is related with speaker diarization and again there are several option to solve it. Kaldi has some good implementations:

https://kaldi-asr.org/