Scaling Armenian ASR: A Semi-Automated Approach to Building High-Performance Speech Recognition Systems for Low-Resource Languages

This paper presents an approach to building automatic speech recognition (ASR) for Armenian, a lowresource language.We created a semi-automated pipeline that collected ~8,000 hours of speech from YouTube and other public sources, validating ~5,350 hours.Transcriptions were generated with a base ASR model and refined with a statistical language model trained on 17M text lines, followed by manual filtering.We trained two architectures adapted from NVIDIA NeMo: a lightweight Streaming Citrinet for low-resource settings and a higher-capacity FastConformer-CTC.Audio preprocessing included noise cancellation and VAD-based chunking for long recordings and implicit language detection.A multilingual pipeline with CNN+RNN-based language ID further enabled transcription across six languages, including Armenian.Our system achieved a best WER of 8.56%, contributing both improved ASR performance and new resources for Armenian NLP, with a framework applicable to other low-resource languages.