Google AI Launches Universal Speech Model (USM): Aims for 1,000 Languages
With the project, Google aims to provide support for the most-spoken languages in the world.
Google has launched its latest update to its Universal Speech Model (USM) to compete with the rapidly-growing popularity of Open AI. The project can currently cover around 300 languages, while the tech giant aims to bolster its capabilities to 1,000 languages. According to Google, the USM surpasses Open AI in all aspects of automated speech recognition operations.
USM comprises speech models with over 2 billion parameters, with a staggeringly large dataset covering around 300 languages with approximately 28 billion sentences, covering 12 million hours of speech. The model is currently used in YouTube, including some relatively uncommon languages, such as Assamese and Cebuano, to generate live captions.
See More: Microsoft Introduces AI-Based Dynamics 365 Copilot
Project Moves Towards Fine-Tuning Self-Supervised Learning
The USM is designed with standardized encoding-decoding architecture, where the encoder uses convolutional, attention, and feed-forward modules, which are compatible with LAS, CTC, and RNN-T decoders.
In their efforts to meet the 1000-language objective, Google is using a machine learning model with the most commonly used languages. Many languages, however, have few actual speakers, limiting data. Google aims to mitigate this problem by collecting and analyzing data from multiple language sources to generate more accurate results through automatic speech recognition scaling.
The training pipeline of the machine learning model comprises self-supervised learning of audio files comprising several languages, followed by optimization of quality and language coverage through text data depending on availability. This is then followed by the fine-tuning of downstream operations with the aid of supervised data.
While the data from Youtube is limited compared to that of OpenAI, the USM model has generated 30% lesser word errors than Open AI. The development of USM is a vital part of the tech giant’s mission to make information organized and universally accessible.
Let us know if you enjoyed reading this news on LinkedIn, Twitter, or Facebook. We would love to hear from you!
Image source: Shutterstock
MORE ON ARTIFICIAL INTELLIGENCE
