Uzbek Speech-to-Text Evaluation Set with 745 Telegram Voice Messages

745 audio files totaling 1 hour and 40 minutes of Uzbek conversational speech, collected from open Telegram groups. The dataset was created by OvozifyLabs for evaluating speech-to-text models and was last updated on December 10, 2025. It features natural voice messages recorded in diverse acoustic conditions and speaking styles.

Use Cases

• Benchmarking Uzbek STT model accuracy based on real-world conversational speech data.
• Evaluating model robustness based on diverse acoustic conditions mentioned in the description.
• Testing STT performance on natural speaking styles captured from voice messages.

Strengths

• 745 audio samples provide a substantial test set for evaluation.
• Total duration of 1 hour 40 minutes (~100 minutes) offers a realistic volume of speech.
• Average duration of ~8 seconds per sample suggests short, focused utterances suitable for testing.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Data may reflect source bias inherent to Telegram group participants.

Provenance

Source

OvozifyLabs

Collection Method

Collected from various open Telegram groups.

Time Range

null

Freshness

Last updated 2025-12-10 13:19:24; freshness should be verified.

Geography

null