Real-Time Automatic Continuous Speech Recognition System for Kannada Language/Dialects

In this work, we present recent advancements in our earlier automatic continuous Kannada speech recognition (ACKSR) system under real-time conditions. In our previous research, we collected task-specific Kannada speech data from 2400 speakers in field conditions, proposing a robust noise elimination technique to enhance degraded speech data. The automatic speech recognition models were developed using Kaldi, and experimental results revealed slightly higher word error rates, attributed to the substantial speech data required for training deep neural networks. Building upon these findings, our current work addresses this limitation by expanding the database. We collected continuous Kannada speech data from an additional 300 speakers under real-time conditions. The updated degraded speech database underwent enhancement using the proposed noise elimination technique. The results demonstrate a significant improvement in the performance of the ACKSR system, particularly in terms of speech recognition accuracy compared to our earlier work.