Conformer2

Conformer-2 Overview

Conformer-2 is an advanced generative AI application designed to leverage spoken data for accurate speech to text conversion. It is a critical component for product and development teams building AI pipelines. The model has been improved from its predecessor, Conformer-1, by utilizing techniques such as model ensembling and scaling up to 1M+ hours of data and model parameter scaling. Conformer-2 also offers significant speed improvements, being up to 55% faster than Conformer-1 depending on the duration of the audio file. It has been designed to improve performance for domains relevant to real-world use cases, with significant improvements in Alphanumeric Transcription Accuracy, Proper Noun Error, and Robustness to Noise.

Conformer-2 Highlights

• Conformer-2 utilizes model ensembling, a technique that uses multiple strong teacher models to produce labels, resulting in a more robust model when exposed to unseen data.
• The model has been scaled up to handle 1.1 million hours of audio data, with a model size of 450M parameters, resulting in improved performance.
• Conformer-2 offers significant speed improvements, being up to 55% faster than its predecessor, Conformer-1.
• The model shows improved performance in Alphanumeric Transcription Accuracy, Proper Noun Error, and Robustness to Noise, making it more suitable for real-world use cases.