Recognition of phonemes in A-cappella recordings using temporal patterns and mel frequency cepstral coefficients

In this paper, a new method for recognizing phonemes in singing is proposed. Recognizing phonemes in singing is a task that has not yet matured to a standardized method, in comparison to regular speech recognition. The standard methods for regular speech recognition have already been evaluated on vocal records, but their performances are lower compared to regular speech. In this paper, two alternative classification methods dealing with this issue are proposed. One uses Mel-Frequency Cepstral Coefficient features, while another uses Temporal Patterns. They are combined to create a new type of classifier which produces a better performance than the two separate classifiers. The classifications are done with US English songs. The preliminary result is a phoneme recall rate of 48.01% in average of all audio frames within a song.