Blind Source Separation of Speech and Background Music for Improved Speech Recognition

Consider the setup where one wants to apply ASR in the presence of background music, as is often the case with broadcast news or documentary programmes. The speech of these is of good quality, a human has no problem understanding it, but an ASR system (although very robust against moderate white noise) fails terribly when the noise is more structured, since music is disturbing the frequency spectrum in a selective way. Typical audio fragments of this type can be found at e.g. the news website of the BBC (http://news.bbc.co.uk/cbbcnews/) or the news flashes on VRT radio (see http://www.radio1.be) or NOS (see http://omroep.nl/nieuws), but also any documentary programme will add background music to most, if not all, commentary, since this gives the programme a “juicier” flavour. The type of music used for this purpose is such that it does not disturb human understanding of the spoken content and does not distract attention. Hence often the music consists of relatively long constant frequency tones, as can be seen on the spectrogram in Figure 1. In the example of Figure 1, an audio fragment of 31 seconds from a BBC documentary was sampled at 16 kHz and the FFT spectrum was calculated on 512 samples