Data · a Chuvash speech corpus

The training corpus the three models grew on, plus additional corpora labeled automatically.

Training corpus · ChuvashAsrDataset

36,781 seg. · ≈49.5 h

Labeled speech from three open sources — the data all three models were trained on. The validation and test sets consist of Common Voice recordings only.

alexantonov / chuvash_voiceread speech · bulk of the training setCC0
38.7 h29,860 seg.
Common Voice 25.0 (cv)part of training + all validation and testCC0
5.6 h3,986 seg.
ftyers · Turkic_TTSChuvash TTS corpusCC-BY-SA
5.2 h2,935 seg.
Labeled total
49.5 h36,781 seg.
train 34,251 · 45.9 h · three sourcesval 1,242 · 1.67 h · Common Voice onlytest 1,288 · 1.91 h · Common Voice only

Additional corpora · labeled by Aisar-turbo

687,322 seg. · ≈1,080 h

Audio from open sources and crowdsourcing: recordings are split into short segments and transcribed automatically by the Aisar-turbo model. None of this data was used to train the models.

ChuvashRawaudiobooks · Bible · lyricsCC BY-NC 4.0
219.0 h370,976 seg.
ChuvashConversationRawconversations · spontaneous speechCC BY-NC 4.0
125.3 h61,020 seg.
ChuvashNewsRawnews · broadcast · diverse speechCC BY-NC 4.0
735.8 h255,326 seg.
Total
1080.1 h687,322 seg.

How the corpus was built

01 · collection

Open sources and crowdsourcing

Audiobooks, the Bible, songs, everyday conversations and news broadcasts — a range of speakers, genres and recording conditions.

02 · segmentation

Silero VAD

Long recordings are split at pauses into fragments of 0.5–60 seconds and stored as FLAC at the original sample rate.

03 · labeling

Transcribed by the model

The texts of the additional corpora were generated by Aisar-turbo, trained in this work. None of that data went into model training.

Note. The transcriptions of the additional corpora are model-generated and contain errors. The recordings include Russian speech, so for Chuvash-only tasks the data should be filtered further. A check on ChuvashRaw (MMS-LID plus a binary Russian/Chuvash classifier) showed that of ≈297 analyzed hours, ≈256 hours (≈86%) are predominantly Chuvash speech (score_chv ≥ 0.5 threshold).
The three additional corpora are CC BY-NC 4.0: free for research and other non-commercial use; commercial use by agreement with the SpeechCollector community (Telegram). Please do not use the data for surveillance of individuals. ChuvashAsrDataset keeps the licenses of its source corpora (CC0 / CC-BY-SA).