Building stuff with machine learning and natural language processing.
Since the arrival of the Transformer architecture in 2017, transformer-based models have made their way into all domains of machine learning. As Andrej Karpathy of Tesla noted, they are quickly becoming the go-to architecture as either a strong baseline or state-of-the-art performance for most problems in language, vision, and now speech.
It’s an exciting time to be working in speech processing. The field has seen a lot of transformations and changes just over the past few years. In this post, I’d like to share my perspective on how I see the field moving forward in 2022, based on what we saw recently.
Research in self-supervised learning for speech has been accelerating since the original wav2vec model released by Facebook AI in 2019. HuBERT is one of the latest of such models, with an open-source implementation already available in HuggingFace’s Transformers library.
Transformer-based neural networks have been revolutionizing the natural language processing field, but are only starting to become popular in the speech processing community. Wav2vec 2.0 is set to change that. Its architecture is based on the Transformer’s encoder, with a training objective similar to BERT’s masked language modeling objective, but adapted for speech.
No matter if you are training a model for automatic speech recognition or something more esoteric like recognizing birds from sound, you could benefit a lot from audio data augmentation. The idea is simple: by applying random transformations to your training examples, you can generate new examples for free and make your training dataset bigger.
Over the past decade, there has been tremendous progress in using artificial intelligence to better understand people. From natural language understanding to recognizing people’s emotions, deep learning has enabled a whole new set of applications that will profoundly change the way we interact with machines.
Wav2vec is a speech encoder model released by the Facebook AI team in late 2019. It quickly became popular in the speech processing community as it enabled new state-of-the-art performance for various speech tasks like automatic speech recognition or emotion recognition. Since its original version, the model has been upgraded to use the Transformers architecture (the same one behind GPT and BERT)…
Since the deep learning wave started in the early 2010s, there has been much hype and disappointments. I feel that a big part of this is due to high expectations driven by research progress that do not translate so well in real-world applications. Hopefully, self-supervised learning might be able to close the gap between these two worlds. This learning paradigm is not new, but it has seen a…
Full paper available on arXiv . Speech emotion recognition has become more and more popular over recent years, largely because of the large range of applications enabled by the technology in fields ranging from human-computer interaction to healthcare. Tech giants have also seen this future potential and have been launching new products such as the Amazon Halo which can detect emotions in…
Currently living in Taiwan, I recently joined the Taiwan Bot 🤖 project along with Shawn and Eric . The idea is to build a go-to assistant to help foreigners answer their questions about moving to, working, and living in Taiwan (pro-tip: ask the bot where to find cheese or chocolate). Building a functional and useful chatbot is a non-trivial project. Fortunately, there has been impressive progress…