Meta’s MMS – Speech2Text & Text2Speech

Dig Deeper

Meta annoncerer massive fremskridt inden for sproggenkendelse og talesyntese

Meta, tidligere kendt som Facebook, har gjort store fremskridt inden for AI taleteknologi. Projektet kaldet Massive Multilingual Speech (MMS), har nu skaleret tale til tekst og tekst til tale for over 1.100 sprog, en ti gange stigning fra tidligere arbejde. MMS er en spilændrings teknologi, da den deler visionen om at bryde sprogbarrierer ved at gøre det muligt for maskiner at genkende og producere tale i over 1.000 sprog og dermed dække 10 gange flere sprog end nogen eksisterende talegenkendelses- eller talegenereringsmodel.

Udfordringer ved at skabe multilinguale systemer for talegenkendelse og taleproduktion

En af de største udfordringer ved at skabe multilinguale systemer til talegenkendelse og taleproduktion er at finde de store mængder labeldata, der er nødvendige for at træne gode modeller til taleopgaver. I de fleste sprog mangler der simpelthen sådanne data, hvilket gør det næsten umuligt at udvikle kvalitetsmodeller til taleopgaver.

MMS-projektet bruger bibeloversættelser til at skabe datasæt i over 1.100 sprog

MMS-projektet overvinder disse udfordringer ved at bruge bibeloversættelser, der har lydoptagelser af mennesker, der læser teksten på forskellige sprog, til at oprette datasæt med oplæsninger på over 1.100 sprog. Dette kombineret med unlabelledt lydoptagelser fra forskellige kristne religiøse oplæsninger giver labeldata for 1.100 sprog og unlabelledt data for næsten 4.000 sprog.

MMS-model udperformes eksisterende modeller og har lige god ydelse på mandlige og kvindelige stemmer

Resultaterne af dette projekt viser, at MMS-modeller overgår eksisterende modeller og dækker 10 gange så mange sprog. Derudover fungerer deres modeller lige godt for mandlige og kvindelige stemmer.

Stor indvirkning på kundesupport og kommunikation på tværs af sprog

MMS-teknologien kan have stor indvirkning på kundesupport og kommunikation på tværs af sprog. Ved at gøre det muligt for mennesker at henvende sig på deres foretrukne sprog, vil det ikke blot forbedre kundeoplevelsen, men også hjælpe virksomheder med at nå et bredere globalt publikum og føre til mere effektiv kommunikation på tværs af forskellige sprog og kulturer.

Video beskrivelse

Well, here is some super exciting news from yesterday! Meta's Massively Multilingual Speech project has made impressive progress in enabling machines to recognize and produce speech in over 1,100 languages, covering ten times more languages than any existing speech recognition or speech generation model. It's heartening to see that Meta is working towards breaking down language barriers and enabling people from all over the world to access information and communicate effectively in their preferred language.

The MMS project has addressed a significant challenge in creating a multi-modal system for speech recognition and generation by using Bible translations to create a dataset of readings in over 1,100 languages, combined with unlabeled audio recordings from various religious readings, providing labeled data for 1,100 languages and unlabeled data for nearly 4,000 languages.

What's exciting is that Meta has provided code and models publically so that other researchers can build upon their work and make contributions to the preservation of language diversity. This is an important step towards creating a world where technology can understand and communicate in any language, effectively preserving linguistic diversity. Overall, the Massively Multilingual Speech (MMS) project represents a significant step towards achieving this goal.

The Massively Multilingual Speech (MMS) project is an incredible advancement in the field of artificial intelligence and language technology. The project's ability to recognize and produce speech in over 1,100 languages is revolutionary and has the potential to make a significant impact on how people communicate across the world.

One of the most significant challenges in creating a multi-model system for speech recognition and speech generation is finding the vast amounts of labeled data necessary for training such models. The MMS project overcomes this challenge by using the New Testament translations, providing labeled data for 1,100 languages and unlabeled data for nearly 4,000 languages. The results of this project show that the MMMS models outperform existing models and cover ten times as many languages.

This project has the power to make a significant impact on the future of communication, breaking down the language barriers that have historically existed and encouraging people to preserve their languages. The ability to communicate effectively in any language will positively affect global commerce and cultural exchange.

Moreover, the MMS project provides code and models publicly so that other researchers can build upon their work and make contributions to the preservation of language diversity in the world. The project recognizes the importance of collaboration within the AI community to ensure that AI technologies are developed responsibly, which is crucial for promoting trust and avoiding unintended consequences.

The MMS project represents a significant step towards achieving the goal of creating a world where technology can understand and communicate in any language, effectively preserving linguistic diversity. It holds the potential to change the way the world thinks about cross-linguistic communication. With the advancements of technology, the future is looking bright for communication across languages and cultures.

In conclusion, the MMS project signifies a fantastic achievement in the field of artificial intelligence and language technology. It will undoubtedly impact the future by breaking down language barriers and promoting linguistic diversity. It is exciting to think about the potential positive implications of this project and other advancements in AI technology in facilitating an inclusive global conversation.

Source:https://ai.facebook.com/blog/multilingual-model-speech-recognition/?utm_source=twitter&utm_medium=organic_social&utm_campaign=blog&utm_content=card

Source: Github: https://github.com/facebookresearch/fairseq/tree/main/examples/mms

Source Meta Video: https://twitter.com/i/status/1660722199395704852

Source: FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech
https://arxiv.org/abs/2205.12446

Source: Robust Speech Recognition via Large-Scale Weak Supervision: https://arxiv.org/abs/2212.04356