Hinotsuba/sherpa-onnx-omnilingual-asr-1600-languages-300M-ctc-int8-2025-11-12

Omnilingual ASR — Local Model Bundle

This repository provides model assets from Meta AI's Omnilingual ASR project for convenient local use and experimentation.

Omnilingual ASR is a multilingual automatic speech recognition system designed to support more than 1,600 languages, including many languages with limited or previously unavailable ASR support.

This is an independent repository and is not an official Meta AI repository. The original model architectures, weights, training methods, datasets, and research were developed and released by Meta AI and the Omnilingual ASR Team.

About Omnilingual ASR

The original project provides several model families for multilingual speech recognition, including:

• —Wav2Vec2-based models for speech representation learning.
• —CTC models for efficient automatic speech recognition.
• —LLM-based ASR models with optional language conditioning.
• —Zero-shot models designed to generalize to languages with limited training data.

Model sizes in the original release range from approximately 300M to 7B parameters, depending on the architecture.

Intended Use

These models are intended for research, experimentation, and development of multilingual speech recognition systems.

Typical use cases include:

• —Local Speech-to-Text applications.
• —Multilingual transcription.
• —Language-specific ASR pipelines.
• —Research involving low-resource languages.
• —Integration into local AI assistants and speech-processing tools.

Runtime requirements depend on the specific model and inference implementation being used.

Upstream Project

The original Omnilingual ASR project, documentation, inference code, supported-language information, datasets, and research materials are maintained by Meta AI / the Omnilingual ASR Team.

Original project:

Meta AI — Omnilingual ASR: Open-Source Multilingual Speech Recognition for 1600+ Languages

Hugging Face organization:

facebook / Meta AI

For implementation details, training recipes, supported languages, and official model documentation, refer to the original Omnilingual ASR project and its official Hugging Face repositories.

License & Attribution

The original Omnilingual ASR code and model weights are released under the Apache License 2.0.

This repository does not claim authorship of the original model architectures, model weights, datasets, or research.

All applicable copyright, attribution, and license terms from the original Omnilingual ASR release remain in effect.

Original Work

Omnilingual ASR Team et al. Omnilingual ASR: Open-Source Multilingual Speech Recognition for 1600+ Languages.

The original project and its authors should be cited when the models are used in research or derived academic work.

This repository exists primarily as a convenient location for local deployment and experimentation with Omnilingual ASR assets.