Short Utterances based real time on-Device Spoken Language Identification

Speech based tasks on mobile devices such as voice based interactions with virtual personal assistants, speech translation etc. have become commonplace today. For such tasks, real-time language identification can serve as an important first step. However, due to the less speech content in short utterances, this is a challenging problem. In this paper, we present Short Utterances Language Identification (SULInet) which is an attention-based light-weight CNN model. It uses 3 seconds utterance and is fit for real time analysis for resource constraint environment like mobile devices. We used state of the art systems trained on Voxforge and EuroSpeech datasets as baseline. We showcase the results at frame-level (3 seconds) and file-level (complete audio file) on PC and an Android based mobile device. We significantly improved our accuracy and achieved SOTA results by proposing a voice activity detector (VAD) as a preprocessing step. On Voxforge dataset, we surpassed SOTA system by 2.36% with 65% less training parameters and 70% smaller input. Similarly, on EuroSpeech dataset, we achieved comparable results with 42% less parameters. Our approach achieved on-Device frame level accuracy as 92%, 80% on Voxforge and EuroSpeech respectively with just one-tenth of PC model size.