Pioneering the next generation of Voice AI technology

Listen to real clips where DER ranks a flawless speaker-attributed transcript below a broken one, and why cpWER and speaker count error are the metrics to trust.

We are excited to introduce Universal-2, our latest and most powerful Speech-to-Text model trained on over 12.5 million hours of multilingual audio data. Universal-2 achieves substantial improvements over Universal-1 by addressing key last-mile challenges in contemporary Speech-to-Text to deliver highly accurate and readable transcripts for best-in-class results.

Universal-2-TF introduces a two-stage neural text formatting model for ASR that combines token classification and sequence-to-sequence approaches to efficiently handle punctuation, capitalization, and text normalization while achieving superior accuracy across diverse domains.

We are excited to introduce Universal-1, our latest and most powerful speech recognition model. Trained on over 12.5 million hours of multilingual audio data, Universal-1 achieves best-in-class speech-to-text accuracy across four major languages: English, Spanish, French, and German.

We're introducing Conformer-2, our latest AI model for automatic speech recognition. Conformer-2 is trained on 1.1M hours of English audio data, extending Conformer-1 to provide improvements on proper nouns, alphanumerics, and robustness to noise.

An evaluation protocol for unsupervised RL representations uses two linear probing tasks to predict rewards and expert actions, reducing computational cost and improving RL training efficiency.

We're introducing Conformer-1, a state-of-the-art speech recognition model trained on 650K hours of audio data that achieves near human-level performance and robustness across a variety of data.