Human beatbox sound recognition using an automatic speech recognition toolkit

Human beatboxing is a vocal art making use of speech organs to produce drum sounds and imitate musical instruments. Beatbox sound classification current challenge that can be used for automatic database annotation music-information retrieval. In this study, large-vocabulary human-beatbox recognition system was developed with an adaptation Kaldi toolbox, widely-used tool recognition. The corpus consisted eighty boxemes, which were recorded repeatedly by two beatboxers. annotated transcribed the means beatbox specific morphographic writing (Vocal Grammatics). recognition-system robustness recording conditions assessed on recordings six different microphones settings. decoding part made monophone acoustic models trained classical HMM-GMM model. A change features (MFCC, PLP, Fbank) variation parameters tested: (i) number HMM states, (ii) MFCC, (iii) presence or not pause boxeme in right left contexts lexicon (iv) rate silence probability. Our best model obtained addition each lexicon, 0.8 probability, 22 MFCC three states HMM. Boxeme error such configuration lowered 13.65%, 8.6 boxemes over 10 well recognized. settings did greatly affect performance, apart from closed-cup technique.