NE-SpeechEmbed

First speech-text retrieval model for Northeast Indian languages.

Architecture

• Speech encoder: Whisper-medium (fine-tuned on NE-ASR, checkpoint-8000)
• Text encoder: xlm-roberta-base
• Projection: Linear(1024→768) speech, Linear(768→768) text
• Loss: InfoNCE with learned temperature
• Embed dim: 768 (L2 normalized)

Languages

Khasi, Garo, Mizo, Nagamese, Kokborok, Assamese, Wancho, Chakma

Retrieval Results (100-sample pool)

Language	R@1	R@5	R@10
Khasi	5.0%	19.0%	27.0%
Garo	7.0%	21.0%	33.0%
Mizo	5.0%	17.0%	30.0%
Nagamese	5.0%	19.0%	30.0%
Kokborok	8.0%	24.0%	35.0%
Random	1.0%	5.0%	10.0%

Training Data

73,476 speech-text pairs from proprietary MWire corpora and Vaani.

Citation

If you use this model, please cite:

@inproceedings{nyalang-borah-2026-ne,
    title = "{NE}-{M}ulti{S}peech: A Multilingual Speech Corpus and {ASR} Benchmark for {N}ortheast {I}ndian Languages",
    author = "Nyalang, Badal  and
      Borah, Angana",
    booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
    month = oct,
    year = "2026",
    address = "Budapest, Hungary",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.emnlp-main.2471/",
    pages = "56109--56119",
    ISBN = "979-8-89176-459-0"
}

License

CC-BY-4.0 — MWire Labs