Conformer-2

The AssemblyAI team has released Conformer-2, a new model for speech recognition. This updated version of the company's original Conformer model is designed to offer improved accuracy and noise robustness. According to the company, Conformer-2 showed a 30.7% relative reduction in mean character error rate (CER) on their newly curated alphanumeric dataset compared to the original Conformer model. In addition, Conformer-2 demonstrated increased noise robustness when tested against added white noise at various signal-to-noise ratios (SNRs). The updated model was trained on in-house hardware using a fault-tolerant and highly scalable Slurm scheduler. The launch of Conformer-2 also brings a new speech_threshold API parameter, which allows users to set a threshold for the proportion of speech that must be present in an audio file for it to be processed.