Meta's single model translates 100 languages across speech and text

SeamlessM4T is an all-in-one multilingual AI model that handles both speech and text translation within a single architecture.

Traditional speech translation systems usually stitch together multiple separate models: one to transcribe audio into text, another to translate the text, and a third to synthesize the spoken output. Each handoff adds latency and compounds translation errors.

SeamlessM4T replaces that multi-step pipeline with a single unified system, significantly cutting down errors and processing delays across languages.

What it does

SeamlessM4T supports nearly 100 languages across five core multimodal and text tasks.

For listening and transcribing, the model provides automatic speech recognition for nearly 100 languages, alongside speech-to-text translation across nearly 100 input and output languages.

For real-time voice translation, it handles speech-to-speech translation from nearly 100 input languages into 36 output languages, including English.

For text tasks, it covers text-to-text translation across nearly 100 languages, and can also take nearly 100 languages of text and turn them into synthesized speech in 35 output languages.

How it was built and what is open

The model combines findings from several of Meta's earlier breakthrough projects.

It incorporates technology from No Language Left Behind, the 200-language text translation model currently integrated into Wikipedia. It also builds on the direct speech-to-speech translator demo previously developed for Hokkien, as well as the Massively Multilingual Speech project, which covered speech recognition and synthesis across more than 1,100 languages.

Meta is releasing SeamlessM4T under a research license so outside developers and academics can experiment with it immediately.

Along with the model checkpoints, Meta is sharing metadata for SeamlessAlign. With 270,000 hours of mined speech and text alignments, it marks the largest open multimodal translation dataset available today. Technical breakdowns and links are accessible via Meta's AI blog.

Why it matters

For developers building cross-language voice apps or real-time interpretation tools, SeamlessM4T solves the biggest headache in voice tech: chained pipeline latency. Being able to go directly from speech in one language to speech in another within a single neural network makes seamless communication feel much closer to reality.

Keep in mind that the current research license restricts immediate commercial deployment, and voice output coverage at 36 languages remains smaller than the 100 supported for input. Still, this release sets a new open benchmark for universal translation, proving that unified multimodal models will eventually replace fragmented audio pipelines.