SpeechBrain is an open-source all-in-one speech toolkit based on PyTorch. It is designed to make the research and development of speech technology easier. Alongside with our documentation this tutorial will provide you all the very basic elements needed to start using SpeechBrain for your projects.
Open in Google ColabIn this tutorial, we provide a high-level description of the speech tasks currently supported by SpeechBrain. We also show how to perform inference on speech recognition, speech separation, speaker verification, and other applications.
Open in Google ColabOne key component of deep learning is iterating the dataset multiple times and performing parameter updates. This process is sometimes called the "training loop" and there are usually many stages to this loop. SpeechBrain provides a convenient framework for organizing the training loop, in the form of a class known as the "Brain" class, implemented in speechbrain/core.py. In each recipe, we sub-class this class and override the methods for which the default implementation doesn't do what is required for that particular recipe.
Open in Google ColabAn essential part of any deep learning pipeline is the definition of hyperparameters and other metadata. These data in conjunction with the deep learning algorithms control the various aspects of the pipeline, such as model architecture, training, and decoding. At SpeechBrain, we decided that the distinction between hyperparameters and learning algorithms ought to be evident in the structure of our toolkit, so we split our recipes into two primary files: experiment.py and hyperparams.yaml. The hyperparams.yaml file is in a SpeechBrain-developed format, which we call "HyperPyYAML". We chose to extend YAML since it is a highly readable format for data serialization. By extending an already useful format, we were able to create an expanded definition of hyperparameter, keeping our actual experimental code small and highly readable.
Open in Google ColabSetting up an efficient data loading pipeline is often a tedious task which involves creating the examples, defining your torch.utils.data.Dataset class as well as different data sampling and augmentations strategies. In SpeechBrain we provide efficient abstractions to simplify this time-consuming process without sacrificing flexibility. In fact our data pipeline is built around the Pytorch one.
Open in Google ColabBy checkpointing, we mean saving the model and all the other necessary state information (like optimizer parameters, which epoch and which iteration), at a particular point in time.
Open in Google ColabSpeechBrain provides two different methods to use multiple GPUs. These solutions follow PyTorch standards and allow for intra- or cross-node training. In this tutorial, the use of Data Parallel (DP) and Distributed Data Parallel (DDP) within SpeechBrain are explained.
Open in Google Colab