Conformer-2 is an advanced automatic speech recognition AI model developed as a successor to Conformer-1. It's designed with robust improvements for decoding proper nouns, alphanumerics, and exhibiting superior performance in noisy environments. This has been achieved through intensive training on a large corpus of English audio data. An advantage of Conformer-2 is that it does not compromise on word error rate compared to Conformer-1, while providing enhanced user-oriented metrics. Further improvements to Conformer-2, in comparison to its predecessor, were realized by augmenting the training data volume and increasing pseudo-label models. Furthermore, with modifications to the inference pipeline, the latency period of Conformer-2 is reduced, thus expediting overall performance. Another critical step-up with Conformer-2 pertains to its innovative training technique that leverages model ensembling. Instead of deriving labels solely from a single 'teacher', labels are generated in this model from multiple 'teachers', ensuring a more versatile and robust model. This has the effect of reducing the impact of individual model failures. The development of Conformer-2 also involved an exploration into data and model parameter scaling, increasing the model size, and extending the training audio data. These approaches were aimed at matching the underutilized potential identified by the 'Chinchilla' paper for large language models. With these updates, Conformer-2 provides faster response times than Conformer-1, bucking the trend of larger models being slower and more expensive.
