Recent Latvian Speech Corpora for Linguistic Research and Technology Development

This paper presents newly created Latvian speech corpora aimed at advancing both linguistic research and speech technology development. Although multilingual models like XLS-R and Whisper have reduced the amount of data needed for fine-tuning speech recognition models even for less-resourced languages, diverse and curated speech corpora remain essential. We provide an overview of several recent Latvian speech corpora, emphasizing their importance for both general-purpose and domain-specific use cases and comparing their design with previously created speech datasets for Latvian. We also introduce a common platform for analysing open-access Latvian speech corpora, and discuss initial evaluation and integration of speech recognition models fine-tuned on the new datasets for practical speech transcription and post-editing applications in research and industry. Finally, we present a competitive open-source speech recognition model for Latvian.