<p>This paper addresses the problem of automatic speech recognition (ASR) for Armenian dialects, an under-resourced and multi-variational language setting. The objective of this study is to develop dialectal ASR models for Armenian and to systematically evaluate cross-dialectal knowledge transfer under low-resource conditions. We introduce a pseudo-labeling and forced-alignment pipeline that enables the construction of high-quality training data from raw speech, resulting in a 70-hour aligned corpus derived from nearly 230 hours of dialectal audio covering Modern Western Armenian and the Artsakh, Lori, and Mush dialects. Building on this dataset, we design and compare three fine-tuning strategies for state-of-the-art multilingual ASR models (Whisper Large v2, Whisper Large v3, and SeamlessM4T v2): (a) dialect-specific fine-tuning, (b) joint multi-dialect fine-tuning, and (3) a two-stage fine-tuning. Our results demonstrate that the two-stage fine-tuning strategy consistently outperforms both single-dialect and joint multi-dialect training, yielding the lowest error rates across all varieties. In particular, Whisper Large v3 achieves the best overall performance with an average WER of 15.75% and CER of 5.33% across all dialects, substantially outperforming existing Armenian ASR systems and commercial APIs. Furthermore, we provide empirical evidence that cross-dialectal transfer is correlated with linguistic proximity, validating a linguistically motivated hypothesis for data reuse in multi-variational ASR. This work introduces and publicly releases the first open datasets and state-of-the-art ASR models for multiple Armenian dialects.</p>