Adapting Conformer-Based ASR Models for Conversations Over the Phone

There are an increasing number of out-of-the-box automatic speech recognition (ASR) solutions to support new and innovative conversational applications across the enterprise landscape. While their performance is generally satisfactory for many uses, there are real limitations when applying them to conversations over the phone in real time, such as the millions that happen every minute in customer service contact centers around the world. We'll show how fine-tuning an out-of-the-box conformer model on an in-house dataset specific to a region and target use case can significantly improve word error rates to unlock more effective voice AI applications. We'll discuss how this approach of creating these custom ASR models with NVIDIA Riva can drive greater model accuracy and efficiency over standard out-of-the-box solutions.