Conformer2

Presenting Conformer-2, our newest AI model for automatic speech recognition. Trained on 1.1M hours of English audio data, it extends Conformer-1 with enhancements in proper nouns, alphanumerics, and noise robustness. Conformer-2 advances our original Conformer-1 release by boosting both model performance and speed. This update delivers a 31.7% improvement on alphanumerics, a 6.8% improvement on Proper Noun Error Rate, and a 12.0% improvement in robustness to noise. These gains result from expanding training data to 1.1M hours and employing more models for pseudo-labeling data. Conformer-1 set state-of-the-art performance with strong noise robustness, ideal for real-world audio conditions. Conformer-2 matches Conformer-1's word error rate while advancing user-oriented metrics. Since Conformer-1's release, our engineering team has reduced inference pipeline latency by up to 53.7%.