Research on Syllable-Based Language Model in Malay Speech Recognition

Malay language has a relatively standardized syllable structure. This paper converts graphemes into syllable symbols. According to some rules, the syllable structure of a word can be obtained with high accuracy. With the vowels, compound vowels, single consonants and double consonants of Malay, this paper formulates the rules and algorithms for automatic syllabification of Malay words, which can be applied to actual speech recognition systems. And then a syllable-based language model can be obtained through the syllabification of words. Using SRILM, a language model perplexity (PPL) calculation tool, this paper calculated the perplexities of different n-gram language models with different units (characters, phonemes, syllables, words) from a public Malay Twitter corpus. The experimental results show that the syllable-based language model is better than the word-based language model, and the character- or phoneme-based language models are better than the syllable-based language model. However, in most of speech recognition systems, a language model needs to cooperate with an acoustic model and the output result is words. And the speech recognition system with syllable-based can perform as well as the system based on deep learning with character-based language model, while the cost of syllable-based speech recognition system is lower.