Development of Low-Latency and Real-Time Filipino Children Automatic Speech Recognition System using Deep Neural Network

As no studies have been made yet to use real-time speakers in performing the Filipino children automatic speech recognition system, this study aims to use the available children speech corpus to train the system using Deep Neural Network technology and use real-time children speakers to test the proposed low-latency Filipino children automatic speech recognition system with low word error rate. The 6.76 hours of raw wav files from University of the Philippines Digital Signal Processing Laboratory Filipino Children’s Speech Corpus is pre-processed, extracted, and trained to create the planned automatic speech recognition system. The time delay neural network module of Kaldi Toolkit with symmetrical time strides is used to train the acoustic model of the system while 3-gram approach is used to train the language model. As results, the proposed Filipino children automatic speech recognition system has an average latency of 0.5188 seconds while the other cloud-based speech-to-text applications such as Google Speech-To-Text has 0.7217 seconds delay, and Whisper API with 1.0420 seconds of lag. Moreover, the said system is also evaluated with different input devices having the Jabra wired headset with the lowest word error rate of 1.49%. Using the same Jabra wired headset, the accuracy of online speech-to-text applications are also computed which results to 6.17% word error rate for Google Speech-To-Text and 11.28% for Whisper API. Therefore, the proposed Filipino children automatic speech recognition system outperforms other online speech recognition applications by 1.4 times faster than Google Speech-to-Text and two times faster than Whisper API having highest accuracy of 98.51% using real-time children speakers.