Unified Modeling of Multi-Domain Multi-Device ASR Systems

Modern Automatic Speech Recognition (ASR) technology is typically fine-tuned for a targeted domain or application to obtain the best recognition results. This requires training and maintaining a dedicated ASR model for each domain, which increases the overall cost. Moreover, fine-tuned model might not be the most optimal way of sharing knowledge across domains. To address this, we propose a novel unified RNN-T based ASR technology that leverages domain embeddings and attention based mixture of experts architecture. Further, the proposed unified neural architecture allows for sharing of data and parameters seamlessly across domains. Our experiments show that the proposed approach outperforms a carefully fine-tuned domain-specific ASR model, yielding up to 10% relative word error rate (WER) improvement and 30% reduction in overall training cost.