Facebook’s MMS Project Breaks Language Barriers in Speech Recognition

Speech Recognition Technology and the Massively Multilingual Speech Project

Speech recognition technology has experienced tremendous growth in recent years, empowering smart devices, digital assistants, and online software to interact with users in a more natural manner. One challenge that stands in the way of creating universally accessible speech recognition systems is the lack of labeled data for training models across various languages.

In an ambitious attempt to address this issue, Facebook has launched the Massively Multilingual Speech (MMS) project, aiming to cover a wide range of languages spoken across the globe.

The problem of developing models capable of handling the staggering 7,000+ languages in use today is further compounded by the inadequate availability of labeled data for many languages. This issue has prompted researchers to explore self-supervised speech representations, which require minimal labeled data to train speech recognition systems effectively.

Facebook’s MMS project combines the wav2vec 2.0 framework with a new dataset featuring labeled data for over 1,100 languages and unlabeled data for nearly 4,000 languages. By leveraging this dataset, the MMS project seeks to overcome existing limitations and offer support for ten times more languages than previous models. To achieve this substantial coverage, Facebook sourced audio data from religious texts, specifically New Testament readings, providing an average of 32 hours of data per language for more than 1,100 languages.

The MMS project demonstrates remarkable performance compared to other state-of-the-art methods, with nearly equal results for male and female voices. Importantly, research has confirmed that the MMS models do not exhibit undue bias towards religious language. Furthermore, the models employ Connectionist Temporal Classification (CTC) to streamline their performance.

Data preprocessing and alignment play a crucial role in the success of the MMS project. This process includes an efficient forced alignment approach, cross-validation filtering, and alignment model training. Despite the challenge of having only a limited 32 hours of data per language, wav2vec 2.0 has proven effective in training low-resource language models.

Facebook’s MMS project has made significant strides in extending speech recognition technology to thousands of languages worldwide. By focusing on efficient self-supervised learning and leveraging a diverse dataset, the project has the potential to vastly improve communication technology by breaking language barriers. Furthermore, Facebook has made the alignment model publicly available, allowing other researchers to harness this innovation and contribute to the development of new speech datasets, ultimately advancing the field of speech recognition technology.