This paper presents a new Livvi-Karelian corpus, addressing challenges encountered in low-resource language research. The main research goal was to collect and annotate new speech data, as well as to create a transcription dictionary. The corpus includes transcripts from radio broadcasts, featuring samples from 17 speakers (7 males and 10 females). Covering about 4.5 hours of audio recordings, it contains 32037 words, thus being a valuable tool for linguistic research. Among the peculiarities of the presented corpus are instances of code-switching between Livvi-Karelian and Russian. The baseline experiments were carried out with the Kaldi toolkit. Hybrid DNN/HMMs with factorized time-delay neural networks were utilized for acoustic modeling, while trigram and LSTM-based models were used for language modeling. The proposed model allowed achieving the Word Error Rate (WER) of 26%.
Database of Annotations of Karelian Speech (AnKaS) includes timestamps, textual transcriptions, and code-switching marks of Livvi-Karelian radio broadcasts.
The database is represented in JSON format. A separate .json file was created for each speaker. The following keys are used:
The files train.txt, dev.txt, and test.txt contains the lists of phrases for training, fine-tuning, and testing the system respectively, where the first number is the speaker's id, and the second number is the phrase's id.
voc.txt contains list of words from textual transcriptions from AnKaS database with their phonemic representation. Vocabulary includes Karelian words as well as most frequent Russian words. Transcriptions for Russian words were made according to Russian transcribing rules.
| Database Features | Value |
|---|---|
| Number of Speakers | 17 (7 male, 10 female) |
| Total Duration | 4.5h |
| Number of Utterances | 4385 |
| Word Occurrences | 32,037 |
| Unique Words | 9,117 |