The Universal-2 model was just released. This is a next-generation Speech-To-Text model. It's a 660M parameter Conformer RNN-T model built on Universal-1's (already great) performance. Some results: • It's 24% better at recognizing proper nouns • It's 21% improvement in alphanumeric accuracy • 15% enhanced text formatting • 73% of users prefer Universal-2 compared to Universal-1 • More accurate and robust, especially in real-world speech complexity This model solves some of the hardest challenges in making Speech-To-Test reliable For the first time, we can say that WER metrics (Word Error Rate) don't capture the full story about how good this model is. You can go here to try this model: assemblyai.com/playground And here is a blog post with their announcement: assemblyai.com/universal-2 @AssemblyAI is behind Universal-2. I am a long-time collaborator of the company.