An automatic speech recognition system in Odia language using attention mechanism and data augmentation

This paper presents an automatic speech recognition (ASR) system developed for the Indian language Odia. In recent years, deep learning models have been used widely to develop ASR systems in various languages and domains. These models demand huge training resources, primarily annotated continuous speech utterances collected from various speakers. However, sufficient speech corpus is not available in many Indian languages. This paper explores the effectiveness of data augmentation in overcoming data scarcity in the Odia ASR task. The baseline system is developed using BiLSTM and the Seq2Seq framework. Next, a portion of the training data is selected based on phonetic richness, and certain augmentation techniques like pitch alteration and time stretching are applied. The augmented data is used along with the actual training data, and a substantial performance improvement is observed. The effectiveness of the attention mechanism in Odia ASR is also explored. When the system is trained through an attention layer embedded with the baseline BiLSTM model, it outperforms the baseline model and existing Odia ASR systems in the literature.