Minimally balanced corpus for speech recognition

This paper reports the method of collecting minimally balanced corpus for speech recognition. Generally balanced corpora are used for training speech recognition systems. However, these balanced corpora are not optimal. The current paper demonstrates that these corpora can be reduced to a varied degree for various phonemes for developing a minimally balanced corpus. The experiments have been developed on ten speakers´ speech data. Recognition accuracy and amount of training data of phonemes have been analyzed. The result for these speakers shows that different phonemes require a different amount of training data for optimal training.