This paper presents a new method for training sequence-to-sequence models for speech recognition and translation tasks. Instead of the traditional approach of training models on short segments containing only lowercase or partial punctuation and capitalization (PnC) sentences, we propose training on longer utterances that include complete sentences with proper punctuation and capitalization. We achieve this by using the FastConformer architecture which allows training 1 Billion parameter models with sequences up to 60 seconds long with full attention. However, while training with PnC enhances the overall performance, we observed that accuracy plateaus when training on sequences longer than 40 seconds across various evaluation settings. Our proposed method significantly improves punctuation and capitalization accuracy, showing a 25% relative word error rate (WER) improvement on the Earnings-21 and Earnings-22 benchmarks. Additionally, training on longer audio segments increases the overall model accuracy across speech recognition and translation benchmarks. The model weights and training code are open-sourced though NVIDIA NeMo.

It was observed in [6] that such long utterances harm the model convergence. We also note that this Recent advances in speech recognition and translation rely on approach may lead to significant padding in mini-batches, resulting hundreds of thousands of hours of Internet speech data. We argue in wasted computation on non-informative frames. We that state-of-the art accuracy can be reached without relying on present an alternative approach to sampling and batching that web-scale data. Canary - multilingual ASR and speech translation allows us to iterate through data twice as fast, while balancing model, outperforms current state-of-the-art models - Whisper, different languages and data sources better. We further accelerate OWSM, and Seamless-M4T on English, French, Spanish, and the training and inference by adopting a FastConformer [7] architecture German languages, while being trained on an order of magnitude and initializing the encoder from a ASR only pretrained less data than these models. Three key factors enables such dataefficient checkpoint.