Recent progress in Automatic Speech Recognition (ASR) has primarily benefited high-resource standard languages, while dialectal speech remains challenging and underexplored. We present an expanded benchmark for low-resource Modern Greek dialects, covering Aperathiot, Cretan, Lesbian, and Cappadocian, spanning southern, northern, and contact-influenced varieties with varying degrees of divergence from Standard Modern Greek. The benchmark provides dialectal transcriptions in the Greek alphabet, following SMG-based orthographic conventions, while preserving dialectal lexical and morphophonological forms. Using this benchmark, we evaluate state-of-the-art multilingual ASR models in a zero-shot setting and by further fine-tuning per dialect. Zero-shot results reveal a clear performance gradient with dialectal distance from Standard Modern Greek, with best WERs ranging from about 60-70% for southern dialects to over 80% for Lesbian and nearly 97% for Cappadocian. Fine-tuning substantially reduces error rates (up to 47% relative WER improvement), with Cappadocian remaining the most challenging variety (best WER 68.17%). Overall, our results highlight persistent limitations of current pretrained ASR models under dialectal variation and the need for dedicated benchmarks and adaptation strategies.

Pontic Greek is a living but endangered Modern Greek dialect that lacks publicly available AI-oriented speech resources and ASR benchmarks. This work reports on the first systematic inference-only (zero-shot) ASR evaluation on authentic Pontic speech. Progress on Pontic ASR is hindered by two coupled challenges: the scarcity of transcribed speech data and the absence of a standardized orthography, which makes it difficult to create consistent reference transcriptions for evaluation. We address these challenges by releasing a new speech corpus of contemporary Pontic as spoken in Northern Greece, derived from natural conversations and provided with manual, utterance-level, time-aligned transcriptions. To reduce annotator bias and increase practical usability, we collect community evidence on written-form preferences via a small questionnaire and use the observed patterns to guide a consistent Greek-script transcription scheme. We use this corpus to perform inference-only (zero-shot) ASR evaluation, benchmarking four state-of-the-art pretrained speech recognition models under a unified evaluation protocol. Results show that zero-shot recognition remains challenging, establishing baseline figures and underscoring the need for dialect-specific data and adaptation.

Automatic Speech Recognition (ASR) models can aid field linguists by facilitating the creation of text corpora from oral material. Training ASR systems for low-resource languages can be a challenging task not only due to lack of resources but also due to the work required for the preparation of a training dataset. We present a pipeline for data processing and ASR model training for low-resourced languages, based on the language family. As a case study, we collected recordings of Pomak, an endangered South East Slavic language variety spoken in Greece. Using the proposed pipeline, we trained the first Pomak ASR model.

Multilingual speakers are able to switch from one language to the other (“code-switch”) between or within sentences. Because the underlying cognitive mechanisms are not well understood, in this study we use computational cognitive modeling to shed light on the process of code-switching. We employed the Bilingual Dual-path model, a Recurrent Neural Network of bilingual sentence production (Tsoukala et al., 2017), and simulated sentence production in simultaneous Spanish-English bilinguals. Our first goal was to investigate whether the model would code-switch without being exposed to code-switched training input. The model indeed produced code-switches even without any exposure to such input and the patterns of code-switches are in line with earlier linguistic work (Poplack,1980). The second goal of this study was to investigate an auxiliary phrase asymmetry that exists in Spanish-English code-switched production. Using this cognitive model, we examined a possible cause for this asymmetry. To our knowledge, this is the first computational cognitive model that aims to simulate code-switched sentence production.