SpeechCollector/ChuvashConversationRaw

Chuvash Audio Dataset (Unlabeled)

This dataset consists of raw, unlabeled audio recordings primarily in the Chuvash language. It is intended for self-supervised learning, pre-training automatic speech recognition (ASR) models, or audio analysis tasks.

Note: While the primary language is Chuvash, some recordings may contain segments of Russian speech or code-switching, reflecting the bilingual nature of the region.

Dataset Statistics

• Total Files: 61,020
• Total Duration: 125.28 hours
• Total Seconds: 450,989.95

Audio Specifications

• Format: Flac

Length Distribution

The dataset contains a variety of recording lengths. Below is the distribution of file durations:

Usage

This dataset does not contain transcriptions. It is suitable for:

• Self-supervised pre-training (e.g., Wav2Vec 2.0, HuBERT)
• Language identification (Chuvash vs. Russian)
• Voice Activity Detection (VAD) analysis

License

Creative Commons Attribution Non Commercial 4.0

License and acceptable use

• License: CC BY-NC 4.0 — free for research and other non-commercial use.
• Commercial use: possible under a separate agreement with the SpeechCollector community — reach out on Telegram: @Michaelya.
• Surveillance: please do not use this dataset to build systems for surveillance, covert monitoring, or tracking of individuals.

Project website (ru / en / cv): https://chuvash-asr.vercel.app