The rapid advancement of neural speech technologies has significantly improved the quality of automatic speech recognition (ASR) and text-to-speech (TTS) systems for high-resource languages. However, such advances remain unevenly distributed: minority and endangered languages still lack adequate datasets, benchmarks, and reusable pipelines, which hinder both technological inclusion and long-term digitalization. Sardinian, a Romance language spoken in Sardinia (Italy), exemplifies this gap due to inadequate and fragmented speech resources, substantial intra-language variation, and heterogeneous recording conditions. In this work, we extend current Sardinian resources by introducing an integrated framework that jointly supports ASR and TTS under low-resource constraints. Our contribution combines (i) the consolidation of two complementary Sardinian speech corpora with distinct acoustic profiles and dialectalcoverage, (ii) a dedicated preprocessing and forced-alignment pipeline to produce trustworthy audio–text pairs suitable for model training, and (iii) systematic adaptation and evaluation of state-of-the-art models for both ASR and TTS. We conduct an extensive experimental study across multiple training configurations, dialectal settings, and cross-dataset evaluations, quantifying the impact of dataset quality and scale on transcription accuracy and generalization, and characterizing trade-offs between perceptual quality and language conformity in synthesis. Overall, the proposed framework provides a reproducible foundation for advancing Sardinian speech technology and offers methodological guidance that can be transferred to further underrepresented languages facing similar data and variability constraints.