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

Omnilingual ASR is an open-source speech recognition system supporting over 1,600 languages — including hundreds never previously covered by any ASR technology. Designed for broad accessibility, it enables new languages to be added with just a few paired examples without requiring specialized expertise or large datasets. By combining scalable zero-shot learning with a flexible model family, Omnilingual ASR aims to make speech technology more inclusive and adaptable for communities and researchers worldwide.

• Huggingface Demo
• Huggingface Dataset
• Paper
• Blogpost

December 2025 Update

We release two suites of models:

• Checkpoints of improved accuracy (CER) for the CTC and LLM-ASR models compared to our existing LLM-ASR model (omniASR_{CTC,LLM}_{300M,1B,3B,7B}_v2).
• A new variant of the LLM-ASR model that supports decoding on unlimited audio length (omniASR_LLM_Unlimited_{300M,1B,3B,7B}_v2). The unlimited audio length models are briefly described in the architecture overview section. It's accuracy is comparable to limited audio length models, however finetuning recipies for this model are currently not supported.

Documentation

Quick Start

• Installation & Basic Usage - Setup and first transcription
• Inference Pipeline - Comprehensive transcription guide with batch processing, language conditioning, and context examples
• Supported Languages - View the complete list of 1600+ supported languages

Models & Architecture

• Model Specifications - Available models, parameters, and memory requirements
• Architecture Overview - Technical details on W2V, CTC, and LLM model families
• Asset Management - Configuration system for models, tokenizers, and datasets

Training & Data Pipeline

• Data Preparation - End-to-end guide for multilingual dataset preparation, HuggingFace integration, and parquet processing
• Training Recipes - Pre-configured workflows for CTC and LLM model training

---

Installation

The models were developed using fairseq2, a research-focused sequence modeling toolkit. While we provide a reference inference pipeline that works across platforms, audio support requires libsndfile (Mac: brew install libsndfile; Windows may need an additional setup).

# using pip
pip install omnilingual-asr

# using uv
uv add omnilingual-asr

Inference

from omnilingual_asr.models.inference.pipeline import ASRInferencePipeline

pipeline = ASRInferencePipeline(model_card="omniASR_LLM_Unlimited_7B_v2")
audio_files = ["/path/to/eng_audio1.flac", "/path/to/deu_audio2.wav"]
lang = ["eng_Latn", "deu_Latn"]
transcriptions = pipeline.transcribe(audio_files, lang=lang, batch_size=2)

More details on running specific models can be found in the src/omnilingual_asr/models/inference directory.

⚠️ Important: Currently only audio files shorter than 40 seconds are accepted for inference on CTC and LLM model suites.

Supported Languages

To view the full list of 1600+ supported languages, you can access the language list programmatically:

from omnilingual_asr.models.wav2vec2_llama.lang_ids import supported_langs

# Print all supported languages
print(f"Total supported languages: {len(supported_langs)}")
print(supported_langs)

# Check if a specific language is supported
if "eng_Latn" in supported_langs:
    print("English (Latin script) is supported!")

Languages follow the format {language_code}_{script}, for example eng_Latn - English (Latin script), cmn_Hans - Mandarin Chinese (Simplified), ...

Using the HuggingFace Dataset 🤗

We provide a large-scale multilingual speech dataset on HuggingFace under CC-BY-4.0 License: facebook/omnilingual-asr-corpus. This dataset can be directly used with our inference pipeline for evaluation or testing:

pip install "omnilingual-asr[data]"

from datasets import load_dataset
from omnilingual_asr.models.inference.pipeline import ASRInferencePipeline

# Load dataset for a specific language (e.g., Ligurian)
omni_dataset = load_dataset("facebook/omnilingual-asr-corpus", "lij_Latn", split="train", streaming=True)
batch = next(omni_dataset.iter(5))

# Convert to pipeline input format
audio_data = [{"waveform": x["array"], "sample_rate": x["sampling_rate"]}
              for x in batch["audio"]]

# Run inference
pipeline = ASRInferencePipeline(model_card="omniASR_LLM_7B_v2")
transcriptions = pipeline.transcribe(audio_data, batch_size=2)

# Display results
for i, (transcription, original_text) in enumerate(zip(transcriptions, batch["raw_text"]), 1):
    print(f"\n Sample {i}:")
    print(f"   Ground Truth: {original_text}")
    print(f"   Predicted:    {transcription}")

Model Architectures

Model Name	Features	Parameters	Download Size (FP32)	Inference VRAM¹	Real-Time Factor¹ (relative speed)²
omniASR_W2V_300M	SSL	317_390_592	1.2 GiB
omniASR_W2V_1B	SSL	965_514_752	3.6 GiB
omniASR_W2V_3B	SSL	3_064_124_672	12.0 GiB
omniASR_W2V_7B	SSL	6_488_487_168	25.0 GiB
omniASR_CTC_300M	ASR	325_494_996	1.3 GiB	~2 GiB	0.001 (96x)
omniASR_CTC_1B	ASR	975_065_300	3.7 GiB	~3 GiB	0.002 (48x)
omniASR_CTC_3B	ASR	3_080_423_636	12.0 GiB	~8 GiB	0.003 (32x)
omniASR_CTC_7B	ASR	6_504_786_132	25.0 GiB	~15 GiB	0.006 (16x)
omniASR_CTC_300M_v2	ASR	325_494_996	1.3 GiB	~2 GiB	0.001 (96x)
omniASR_CTC_1B_v2	ASR	975_065_300	3.7 GiB	~3 GiB	0.002 (48x)
omniASR_CTC_3B_v2	ASR	3_080_423_636	12.0 GiB	~8 GiB	0.003 (32x)
omniASR_CTC_7B_v2	ASR	6_504_786_132	25.0 GiB	~15 GiB	0.006 (16x)
omniASR_LLM_300M	ASR with optional language conditioning	1_627_603_584	6.1 GiB	~5 GiB	0.090 (~1x)
omniASR_LLM_1B	ASR with optional language conditioning	2_275_710_592	8.5 GiB	~6 GiB	0.091 (~1x)
omniASR_LLM_3B	ASR with optional language conditioning	4_376_679_040	17.0 GiB	~10 GiB	0.093 (~1x)
omniASR_LLM_7B	ASR with optional language conditioning	7_801_041_536	30.0 GiB	~17 GiB	0.092 (~1x)
omniASR_LLM_300M_v2	ASR with optional language conditioning	1_627_603_584	6.1 GiB	~5 GiB	0.090 (~1x)
omniASR_LLM_1B_v2	ASR with optional language conditioning	2_275_710_592	8.5 GiB	~6 GiB	0.091 (~1x)
omniASR_LLM_3B_v2	ASR with optional language conditioning	4_376_679_040	17.0 GiB	~10 GiB	0.093 (~1x)
omniASR_LLM_7B_v2	ASR with optional language conditioning	7_801_041_536	30.0 GiB	~17 GiB	0.092 (~1x)
omniASR_LLM_Unlimited_300M_v2	omniASR_LLM_300M + unlimited audio length	1_627_603_584	6.1 GiB	~5 GiB	0.092 (~1x) (0.206)³
omniASR_LLM_Unlimited_1B_v2	omniASR_LLM_1B + unlimited audio length	2_275_710_592	8.5 GiB	~6 GiB	0.097 (~1x) (0.207)³
omniASR_LLM_Unlimited_3B_v2	omniASR_LLM_3B + unlimited audio length	4_376_679_040	17.0 GiB	~10 GiB	0.095 (~1x) (0.208)³
omniASR_LLM_Unlimited_7B_v2	omniASR_LLM_7B + unlimited audio length	7_801_041_536	30.0 GiB	~17 GiB	0.097 (~1x) (0.208)³
omniASR_LLM_7B_ZS	Zero-Shot ASR	7_810_900_608	30.0 GiB	~20 GiB	0.194 (~0.5x)
omniASR_tokenizer_v1	Tokenizer for all non-v2 models except omniASR_LLM_7B	-	100 KiB	-
omniASR_tokenizer_v1_variant7	Tokenizer for the omniASR_LLM_7B architecture	-	100 KiB	-
omniASR_tokenizer_written_v2	Tokenizer for all v2 architectures	-	100 KiB	-

¹ (batch=1, audio_len=30s, BF16, A100)

² Relative speed to omniASR_LLM_7B

³ (batch=1, audio_len=15min, BF16, A100)

Model Download & Storage

• Automatic Download: Models are automatically downloaded on first use during training or inference
• Storage Location: Models are saved to ~/.cache/fairseq2/assets/

Architecture Documentation

We provide a high-level model architecture overview in the model directory (src/omnilingual_asr/models), with individual configurations for each model family in the respective directories:

• SSL Models: src/omnilingual_asr/models/wav2vec2_ssl
• CTC Models: src/omnilingual_asr/models/wav2vec2_asr
• LLM Models: src/omnilingual_asr/models/wav2vec2_llama

Training

To further finetune the released checkpoints on your own data, use our data preparation guide followed by the finetuning recipe guide.

License

Omnilingual ASR code and models are released under the Apache 2.0.

Citation

If you use the omnilingual ASR model suite in your research and wish to cite us, please use the following BibTeX entry!

@misc{omnilingualasrteam2025omnilingualasropensourcemultilingual,
      title={Omnilingual ASR: Open-Source Multilingual Speech Recognition for 1600+ Languages},
      author={Omnilingual ASR team and Gil Keren and Artyom Kozhevnikov and Yen Meng and Christophe Ropers and Matthew Setzler and Skyler Wang and Ife Adebara and Michael Auli and Can Balioglu and Kevin Chan and Chierh Cheng and Joe Chuang and Caley Droof and Mark Duppenthaler and Paul-Ambroise Duquenne and Alexander Erben and Cynthia Gao and Gabriel Mejia Gonzalez and Kehan Lyu and Sagar Miglani and Vineel Pratap and Kaushik Ram Sadagopan and Safiyyah Saleem and Arina Turkatenko and Albert Ventayol-Boada and Zheng-Xin Yong and Yu-An Chung and Jean Maillard and Rashel Moritz and Alexandre Mourachko and Mary Williamson and Shireen Yates},
      year={2025},
      eprint={2511.09690},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2511.09690},
}