This paper describes AssemblyAI's industrial-scale automatic speech recognition (ASR) system, designed to meet the requirements of large-scale, multilingual ASR serving various application needs. Our system leverages a diverse training dataset comprising unsupervised (12.5M hours), supervised (188k hours), and pseudo-labeled (1.6M hours) data across four languages. We provide a detailed description of our model architecture, consisting of a full-context 600M-parameter Conformer encoder pre-trained with BEST-RQ and an RNN-T decoder fine-tuned jointly with the encoder. Our extensive evaluation demonstrates competitive word error rates (WERs) against larger and more computationally expensive models, such as Whisper large and Canary-1B. Furthermore, our architectural choices yield several key advantages, including an improved code-switching capability, a 5x inference speedup compared to an optimized Whisper baseline, a 30% reduction in hallucination rate on speech data, and a 90% reduction in ambient noise compared to Whisper, along with significantly improved time-stamp accuracy. Throughout this work, we adopt a system-centric approach to analyzing various aspects of fully-fledged ASR models to gain practically relevant insights useful for real-world services operating at scale.

These labels are then used in traditional supervised training schemas. This line of work in turn bifurcates This paper presents Conformer-1, an end-to-end Automatic into two main approaches. The first approach relies on generating Speech Recognition (ASR) model trained on an extensive pseudo-labels using a pre-existing baseline model [1, 6, 7], dataset of 570k hours of speech audio data, 91% of which was while the second approach attempts to source massive amounts acquired from publicly available sources. To achieve this, we of data of ambiguous quality from the public sources and then perform Noisy Student Training [1] after generating pseudolabels filter it down to a subset that is both human labeled and high for the unlabeled public data using a strong Conformer quality [8]. Our work attempts to address the data scarcity issue RNN-T baseline model. The addition of these pseudo-labeled head-on and leverages both data filtering and pseudo-labeling data results in remarkable improvements in relative Word Error to procure high-quality audio and labels at scale. Rate (WER) by 11.5% and 24.3% for our asynchronous and Following the example provided by Whisper [8], we realtime models, respectively. Additionally, the model is more sourced audio speech data from open and fair use sources available robust to background noise owing to the addition of these data.