Automatic speech recognition (ASR) for low-resource languages remains challenging due to limited labeled data. Although multilingual models and the inclusion of related auxiliary languages enable cross-lingual transfer, it is still unclear how introducing cross-lingual information at different training stages-pre-training versus fine-tuning-affects downstream performance. Prior work largely treats transfer as a single-stage optimization problem without disentangling stage effects. We present a stage-aware analysis of cross-lingual transfer for Faroese ASR using related auxiliary languages and Wav2Vec 2.0 XLS-R models. We systematically compare two complementary adaptation pipelines: (i) cross-lingual supervised fine-tuning and (ii) cross-lingual continuous pre-training prior to fine-tuning. Both strategies are evaluated under a unified setup with controlled model architectures, balanced representation of auxiliary languages, and identical evaluation protocols. Results demonstrate that cross-lingual transfer is stage-dependent. Supervised adaptation optimizes in-domain accuracy, while pretraining-level adaptation enhances robustness and reduces Character Error Rate (CER). Auxiliary language effects vary across pipelines, reinforcing the idea that transfer effectiveness depends on when and how cross-lingual information is introduced. Comparisons with large-scale multilingual ASR models highlight trade-offs between model scale and explicit, small-scale domain-aware adaptation. These findings suggest that effective cross-lingual transfer for Faroese low-resource ASR is inherently stage-dependent rather than a single-step design choice.

This work addresses the lack of large-scale, natural speech data for Faroese automatic speech recognition. Existing resources, such as the 100-hour Ravnursson corpus, consist of read speech and do not capture the spontaneous variation, sociolinguistic aspects and prosody of real dialogue, limiting model performance. To overcome this, we present the Faroese Parliament Speech Corpus (FPSC)—a 1,600-hour collection of parliamentary recordings comprising 89,000 speeches with detailed speaker and linguistic metadata. The corpus includes weakly supervised transcriptions generated using an ensemble of four Faroese-adapted ASR models combined through a ROVER-based voting procedure. In creating FPSC, we trained several new state-of-the-art ASR models for Faroese—some built on large-scale pretrained backbones and others leveraging multilingual transfer—all outperforming previously published Faroese ASR systems. FPSC represents the first corpus of natural spoken Faroese and a major step toward realistic ASR modeling for Faroese, offering an open, reproducible, and scalable resource for future speech and language research.

Mozilla Common Voice is a crowdsourced project that aims to create a public, multilingual dataset of voice recordings for training speech recognition models. In Common Voice, anyone can contribute by donating or validating recordings in various languages. However, despite the availability of many recordings in certain languages, a significant percentage remains unvalidated by users. This is the case for Spanish, where in version 17.0 of Common Voice, 75% of the 2,220 hours of recordings are unvalidated. In this work, we used the Whisper recognizer to automatically validate approximately 784 hours of recordings which are more than the 562 hours validated by users. To verify the accuracy of the validation, we developed a speech recognition model based on a version of NVIDIA-NeMo’s Parakeet, which does not have an official Spanish version. Our final model achieved a WER of less than 4% on the test and validation splits of Common Voice 17.0. Both the model and the speech corpus are publicly available on Hugging Face.

ASR development for low-resource languages like Faroese faces significant challenges due to the scarcity of large, diverse datasets. While fine-tuning multilingual models using related languages is a common practice, there is no standardized method for selecting these auxiliary languages, leading to a computationally expensive trial-and-error process. By analyzing Faroese’s positioning among other languages in wav2vec2’s multilingual representation space, we find that Faroese’s closest neighbors are influenced not only by linguistic similarity but also by historical, phonetic, and cultural factors. These findings open new avenues for auxiliary language selection to improve Faroese ASR and underscore the potential value of data-driven factors in ASR fine-tuning.