Conformer

Conformer-2 is a remarkable AI model for automatic speech recognition. Trained on an extensive 1.1M hours of English audio data, it builds upon the success of Conformer-1 and offers significant advancements. One of the key improvements is in handling proper nouns, alphanumerics, and robustness to noise. The model's performance in these areas is truly impressive. For example, it achieves 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 enhancements are made possible through a combination of increased training data and the use of model ensembling. By leveraging multiple strong teachers to generate labels, the student model becomes more robust and better able to handle a wider variety of data. Additionally, the team behind Conformer-2 has made significant improvements in the model's speed. Despite the larger model size, they were able to reduce the latency of the inference pipeline by up to 53.7%, allowing users to get their results faster. Another important aspect of Conformer-2 is its focus on real-world use cases. While traditional metrics like word-error-rate (WER) are useful, the model also takes into account the significance of errors. For instance, the newly crafted Proper Noun Error Rate (PPNER) metric quantifies the model's performance specifically for proper nouns, ensuring more accurate and consistent transcriptions. Conformer-2 was trained on the company's own GPU compute cluster, providing greater flexibility and control over the training process. With the launch of Conformer-2, users also have access to a new API parameter, speech_threshold, which allows them to control the processing of audio files based on the proportion of speech present. Overall, Conformer-2 represents a major advancement in the field of speech recognition and is set to have a significant impact on various applications that rely on accurate speech-to-text conversion.