Meta AI: New Program to Enhance Speech and Translation

Meta and UNESCO Collaborate on Openly Available AI Language Development

Meta is initiating a new collaborative program with UNESCO focused on gathering speech recordings and transcriptions. This data collection effort is intended to facilitate the advancement of openly accessible AI technologies.

The Language Technology Partner Program

The newly established Language Technology Partner Program is actively seeking contributors. These collaborators are expected to provide substantial linguistic resources, including over 10 hours of speech data paired with corresponding transcriptions.

Furthermore, the program requests large volumes of written text and collections of translated sentences, all encompassing a wide variety of languages. Meta plans to integrate these contributions into its AI speech recognition and translation models.

Focus on Underserved Languages

Partners will collaborate directly with Meta’s AI teams. The resulting models, once completed, will be released as open-source resources.

Initial partners include the government of Nunavut, a region in Northern Canada characterized by its low population density. A significant portion of Nunavut’s residents communicate in Inuktut, a group of related Inuit languages.

Meta emphasized its commitment to supporting UNESCO’s objectives, particularly concerning languages that have historically received limited technological support. The company’s stated aim is to develop AI systems capable of understanding and responding to diverse human needs, irrespective of linguistic or cultural origins.

New Machine Translation Benchmark

In addition to the partner program, Meta is releasing an open-source benchmark for evaluating machine translation model performance.

This benchmark, developed by linguists, currently supports seven languages and is available through the Hugging Face AI development platform. Contributions to the benchmark are also welcomed.

Philanthropic Goals and Business Benefits

Meta positions these initiatives as philanthropic endeavors. However, the company is also poised to benefit from improvements in its speech recognition and translation capabilities.

The company is continually expanding the linguistic support for its AI-powered assistant, Meta AI. It is also actively piloting features like automatic translation tools for content creators.

Last September, Meta announced testing of a voice translation feature for Instagram Reels, enabling creators to dub their videos and synchronize lip movements automatically.

Past Criticism of Language Support

Meta’s handling of non-English content has faced scrutiny. A recent report indicated that approximately 70% of COVID-related misinformation in Italian and Spanish remained unflagged, compared to 29% of similar English content.

Internal documents revealed that Arabic-language posts are frequently and incorrectly identified as hate speech.

Commitment to Improvement

Meta has acknowledged these issues and stated its dedication to enhancing its translation and content moderation technologies.