Meta is Focused on Leveraging Religious Texts for AI Training
In Brief
Meta's new AI-enabled speech synthesis technology has the ability to recognize a staggering 4,000 languages.
The primary goal of this initiative is to help ensure the preservation of various languages.
The company is utilizing religious texts, including the Bible, as training material for its advanced multilingual speech models.
Meta has recently introduced a groundbreaking AI text-to-speech tool that can detect over 4,000 languages. This project aims to not only preserve these languages but also enrich communication through the use of well-known religious texts as foundational training data.
"Gathering audio samples across thousands of languages presented an initial hurdle, as existing speech datasets typically include data for only around 100 languages. To address this challenge, we sourced widely translated religious texts like the Bible, which have undergone extensive linguistic research for translation purposes,\" Meta explains in a recent blog post. post .
Meta indicates that a significant portion of their raw data comes from the Bible. They have also obtained recordings and textual materials from platforms such as FaithComesByHearing.com, GoTo.Bible, and Bible.com.
The project has amassed audio data from over 6,255 languages and dialects, incorporating Bible narratives, evangelistic speeches, scripture recitations, and musical pieces. Interestingly, Meta claims that their models are well-suited for both male and female voices, despite the predominance of male voice samples in the recordings.
The data specifically for New Testament readings amounts to about 32 hours of audio available in each language, contributing to a dataset that includes more than 1,100 distinct languages. Advisors from Christian ethics assure that the use of the New Testament and its translations in machine learning does not raise concerns among Christians, and similar principles hold for other sacred texts.
"Although the audio content is religious in nature, our research indicates that this does not skew the model towards generating religious-themed language,\" the blog post articulates.
Thus, the inclusion of religious training materials is not expected to skew the system's output towards any specific ideological stance, nor will it lead to the generation of text that mimics religious prose.
Read more related articles:
- OpenAI Announces Shift in Policy: No Longer Defaulting to Utilize Customer Data for Model Training
- Google Gears Up to Surpass Meta by Launching Imagen Video, a Cutting-Edge Text-to-Video AI Generator
- Meta Reveals ‘Segment Anything,’ an Innovative AI Model for Image Segmentation
Disclaimer
In line with the Trust Project guidelines , please be aware that the content shared on this platform does not constitute legal, tax, investment, or financial advice. It's essential to only invest what you can afford to lose and seek professional financial guidance if you have any uncertainties. For additional clarity, we recommend consulting the terms and conditions, along with support resources provided by the issuer or advertiser. MetaversePost is dedicated to accurate and impartial journalism, although the market landscape can shift unpredictably.
