Description

Conformer-2 is an advanced AI model that has been specifically designed for automatic speech recognition (ASR). It is an upgrade to its predecessor, Conformer-1, and has been trained on an extensive dataset of 1.1 million hours of English audio. The primary focus of Conformer-2 is to enhance the recognition of proper nouns, alphanumerics, and noise robustness, which significantly improves its ability to accurately transcribe spoken content.

Conformer-2 has been developed using the scaling laws proposed in DeepMind's Chinchilla paper, which emphasizes the importance of sufficient training data for large language models. The model leverages a massive 1.1 million hours of English audio data during its training process. One of the standout features of Conformer-2 is its adoption of model ensembling, which reduces variance and enhances the model's performance when dealing with previously unseen data during training.

Despite its increased model size, Conformer-2 exhibits improvements in terms of speed compared to Conformer-1. The serving infrastructure has been meticulously optimized, resulting in faster processing times. Conformer-2 achieves up to a 55% reduction in relative processing duration across all audio file durations.

In real-world applications, Conformer-2 demonstrates significant enhancements in various user-oriented metrics. Notably, it achieves a 31.7% improvement on alphanumerics, a 6.8% improvement on proper noun error rate, and a 12.0% improvement in noise robustness. These enhancements are attributed to both the vast training data and the use of an ensemble of models.

The Conformer-2 model proves to be an invaluable component for AI pipelines that focus on generative AI applications using spoken data. Its remarkable speech-to-text transcription capabilities make it a valuable tool for generating accurate transcriptions with exceptional precision and reliability.