Long Short-Term Memory Based Language Model for Indonesian Spontaneous Speech Recognition

A robust recognition performance in daily or spontaneous conversation becomes necessary for a speech recognizer when deployed in real world applications. Meanwhile, the Indonesian speech recognition system (ASR) still has poor performance compared to dictated speech. In this work, we used deep neural networks approach, focused primarily on using long short-term memory (LSTM) to improve the language model performance as it has been successfully applied to many long context-dependent problems including language modeling. We tried different architectures and parameters to get the optimal combination, including deep LSTMs and LSTM with projection layer (LSTMP). Thereafter, different type of corpus was employed to enrich the language model linguistically. All our LSTM language models achieved significant improvement in terms of perplexity and word error rate (%WER) compared to n-gram as the baseline. The perplexity improvement was up to 50.6% and best WER reduction was 3.61% as evaluated with Triphone GMM- HMM acoustic model. The optimal architecture combination we got is deep LSTMP with L2 regularization.