🇧🇮 The first large-scale, open-source speech and text dataset for Kirundi language. Building AI models for 12M+ Kirundi speakers through community collaboration. Includes ASR, TTS, and MT capabilities.

Experimental general-purpose SIGER LLM built from scratch with SSM architecture, LoRA fine-tuning, evaluation and optimization pipelines, with Lampung Dialek O as an early low-resource language test case.

LLM-Based Evaluation of Low-Resource Machine Translation (CLNLP 2025)

Full-stack AI application for Telugu text summarization using TF-IDF and mT5, with optional speech generation via TTS.

Dialect-aware language model for all six Romansh varieties. QLoRA continued pretraining on ZurichNLP/quotidiana; single-GPU, reproducible. First open Romansh LM.

Com la tokenització en subparaules fractura la morfologia catalana, i si una segmentació morfèmica universal en recupera la geometria. El català es fragmenta ~1,7× més que l'anglès (el punt volat fins a ~4×); auditoria d'11 tokenitzadors i geometria en 5 models petits.

UrduReason-Eval: A comprehensive evaluation dataset with 800 Urdu reasoning problems across 6 categories (arithmetic, logical deduction, temporal, comparative, and causal reasoning) for assessing reasoning capabilities in Urdu language models.

Telugu AI Voice & Text Summarizer 2026 🔥 mT5 & TTS NLP App

FilBench website: filbench.github.io

Final version of Bartangi Lemmatizer and Word2Vec embeddings

This project will explore how well existing ASR systems perform for African languages ( Setswana ) and how extracted transcripts from local language podcasts can be used for topic modeling and how what NLP techniques improve topic modeling.