Singing voice detection in monophonic and polyphonic contexts

In this article, we present an improvement of a previous singing voice detector. This new detector is in two steps. First, we distinguish monophonies from polyphonies. This distinction is based on the fact that the pitch estimated in a monophony is more reliable than the one estimated in a polyphony. We study the short term mean and variance of a confidence indicator; their repartition is modelled with bivariate Weibull distributions. We present a new method to estimate the parameters of these distributions with the moment method. Then, we detect the presence of singing voice. This is done by looking for the presence of vibrato, an oscillation of the fundamental frequency between 4 and 8 Hz. In a monophonic context, we look for vibrato on the pitch. In a polyphonic context, we first make a frequency tracking on the whole spectrogram, and then look for vibrato on each frequency tracks. Results are promising: from a global error rate of 29.7 % (previous method), we fall to a global error rate of 25 %. This means that taking into account the context (monophonic or polyphonic) leads to a relative gain of more than 16 %.