Armenian ASR: State-of-the-Art Speech Recognition

Armenian remains one of the most underserved languages in speech AI. While major languages benefit from massive datasets and highly optimized speech recognition models, Armenian still lacks high-quality open technologies that developers, researchers, and businesses can easily build upon. Our project aims to change that by developing a state-of-the-art Automatic Speech Recognition (ASR) system specifically for Armenian. The project is built on one of the largest gated Armenian speech datasets available, containing more than 8,000 hours of audio from diverse speakers and recording conditions. We are training and evaluating modern speech recognition models while optimizing every stage of the pipeline, including data preprocessing, audio normalization, fine-tuning, decoding, and inference. Beyond model quality, we focus on reproducibility and scalable engineering. Our training pipeline is designed to efficiently leverage modern GPU and TPU infrastructure, allowing us to experiment with different model architectures and optimization strategies. Our long-term vision is to make Armenian one of the best-supported low-resource languages in AI. By improving speech recognition quality and building reliable tooling around it, we hope to accelerate research, enable new AI-powered applications, and strengthen the Armenian AI ecosystem.