Conformer

Description

Conformer. Conformer-2 stands as a state-of-the-art AI model tailored specifically for automatic speech recognition (ASR). Unlike its predecessor, Conformer-1, this advanced model has undergone training on an extensive dataset comprising 1.1 million hours of English audio, resulting in remarkable advancements across various aspects of speech recognition. The primary focus of Conformer-2 lies in enhancing the recognition of proper nouns, alphanumerics, and noise robustness, thereby significantly improving its ability to accurately transcribe spoken content. Drawing inspiration from DeepMind’s Chinchilla paper, the development of Conformer-2 adheres to scaling laws, recognizing the significance of ample training data for large language models. By leveraging a massive 1.1 million hours of English audio data during its training process, Conformer-2 ensures a robust foundation. One of the standout features of Conformer-2 is its implementation of model ensembling, which involves generating labels from multiple strong teachers rather than relying on predictions from a single teacher model. This ensembling technique effectively reduces variance and enhances the model’s performance when confronted with previously unseen data during training. Despite its increased model size, Conformer-2 exhibits improved speed compared to Conformer-1, thanks to meticulous optimization of the serving infrastructure, resulting in faster processing times. In terms of real-world performance, Conformer-2 showcases significant enhancements in various user-oriented metrics. Notably, it achieves a 31.7% improvement in alphanumerics, a 6.8% improvement in proper noun error rate, and a 12.0% improvement in noise robustness. These enhancements can be attributed to both the vast training data and the utilization of an ensemble of models. With its exceptional capabilities, Conformer-2 proves to be an ideal choice for AI pipelines.