Meta has developed AI models capable of recognizing and generating speech for over 1,000 languages, a significant step toward preserving endangered languages. The models are being released as open source on GitHub to facilitate the development of multilingual speech applications.
Existing speech recognition models only cover around 100 out of approximately 7,000 languages globally due to limited labeled training data. Meta overcame this challenge by retraining an AI model on audio recordings of the New Testament in various languages, enabling the algorithm to learn new languages with minimal data.
Meta's models can converse in over 1,000 languages and recognize more than 4,000, surpassing rival models like OpenAI Whisper with a lower error rate. However, the team acknowledges the risk of mistranscribing certain words or phrases and the presence of some biased words.
While the research is impressive, the use of religious texts like the Bible for training AI models raises concerns about bias and misrepresentation. Experts emphasize the need for careful consideration when selecting training data to ensure fairness and accuracy.
Meta's AI models offer great potential for multilingual communication and language preservation. Ongoing efforts to address biases and improve accuracy will be crucial for their responsible and inclusive deployment in various applications.