On-Device Automatic Speech Recognition Application for Android Smartphone using Tensorflow Mobile Library and Bidirectional LSTM

Abstract

An on-device automatic speech recognition application (ASR App) was developed to output a text message from a wav file in an android smartphone. The ASR App uses a Mel-frequency cepstral coefficient (MFCC) based pre-processor, a Bidirectional Long Short Term Memory (BiLSTM) based acoustic model and a trigram based language model. The acoustic model uses the Tensorflow Mobile library. The word error rate (WER) was 7.9% on the Wall Street Journal (WSJ) eval92 language corpus. The average calculation time for 1.0 sec audio input speech was 0.27 sec on Galaxy S9. The on-device ASR App works off-line and is privacy-safe.