PASABI: Pagmensahe ng Salitang Binigkas A Filipino Speech-to-Text Messaging Application Using Recurrent Neural Networks

Abstract

PASABI is a Filipino text messaging mobile application with a speech-to-text functionality. The speech-to-text functionality makes use of Keras models produced with the separate PASABI desktop trainer. The trainer makes use of Recurrent Neural Networks for this task. Connectionist Temporal Classi cation is also utilized by creating a speech-to-text model that is trained by mapping characters in the transcription to the audio. By training the model directly to the characters, the need for speech datasets with phonetic transcriptions, or the development of algorithms to generate these phonetic transcriptions, is removed. The provided trainer can be used to develop models with new data, and be able to deploy it to the mobile application.