Omnilingual SONAR: Cross-Lingual and Cross-Modal Sentence Embeddings Bridging Massively Multilingual Text and Speech

A research team has introduced OmniSONAR, a family of sentence embedding models capable of handling thousands of languages and embedding text, speech, code, and mathematical expressions in a single semantic space, according to a paper posted on arXiv [1]. The model family, detailed in a paper last revised in June 2026, aims to address a persistent limitation in cross-lingual sentence encoders, which typically cover only a few hundred languages and often sacrifice downstream quality for stronger alignment [1][2]. OmniSONAR was built using progressive training. The researchers first learned a foundational space for 200 languages with an LLM-initialized encoder-decoder, combining token-level decoding with a novel split-softmax contrastive loss and synthetic hard negatives [2]. They then expanded to several thousand language varieties through a two-stage teacher-student encoder distillation framework [2]. Finally, they mapped 177 spoken languages into the space to demonstrate cross-modal extensibility [2]. On the 200-language FLORES dataset, OmniSONAR halved cross-lingual similarity search error [1][2]. On the 1,560-language BIBLE benchmark, it reduced error by a factor of 15 [1][2]. The model also outperformed NLLB-3B on multilingual benchmarks and exceeded prior models, including much larger LLMs, by 15 chrF++ points on 1,560 languages into English BIBLE translation [1][2]. For speech, the model achieved a 43% lower similarity-search error and reached 97% of SeamlessM4T speech-to-text quality, despite being zero-shot for translation and trained only on ASR data [1][2]. The paper was authored by João Maria Janeiro and collaborators [1]. The work appears on arXiv, a preprint server that accounts for roughly 95% of the paper URLs Hugging Face users have linked in their repositories, according to Hugging Face documentation [4]. The Hugging Face Hub allows researchers to link papers to models, datasets, and interactive demos, and arXiv has collaborated with Hugging Face to embed demos directly alongside paper abstract pages [4][5]. The OmniSONAR paper is listed with links to code, data, and media, including a Hugging Face entry, on its arXiv page [1]. Large language models have drawn intense investment and competition. Chinese firm DeepSeek, for instance, reported training its V3 model for US$6 million, far less than the reported US$100 million cost for OpenAI's GPT-4 in 2023, and using approximately one-tenth the computing power consumed by Meta's comparable Llama 3.1 model [7]. Alibaba Cloud's Qwen family of LLMs is distributed under open-source licenses including Apache 2.0 [9]. OmniSONAR's authors also trained an encoder-decoder language model, Spectrum, exclusively on English text processing OmniSONAR embedding sequences, unlocking high-performance transfer to thousands of languages and speech for complex downstream tasks [1][2].