Tajik Asr Corpus V3: 1,071 Hours of Tajik Speech for ASR Training

1,071 hours of Tajik speech data form this corpus for automatic speech recognition. It combines machine-labeled audio from 41 YouTube channels with gold-standard transcriptions from the FLEURS benchmark. The dataset was created by Peacockery and was last updated on June 12, 2026.

Use Cases

• Train automatic speech recognition models based on 1,071 hours of Tajik audio.
• Benchmark ASR model performance based on the FLEURS test and conversational speech splits.
• Fine-tune pre-trained speech models based on a large corpus of Tajik speech data.
• Study the characteristics of machine-labeled versus human-annotated speech data for the Tajik language.

Strengths

• Contains 1,071 hours of audio, providing substantial training material.
• Includes 11.8 hours of gold-standard transcriptions from the FLEURS benchmark for reliable evaluation.
• The associated model achieved a 16.9% word error rate on the FLEURS test set, demonstrating the dataset's utility.

Limitations

• The majority of the data (1,059 hours) is machine-labeled, which may introduce labeling noise.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

41 Tajik YouTube channels and the FLEURS tg_tj benchmark.

Collection Method

Audio data was likely gathered from YouTube and processed; labels are a mix of machine-generated and human-annotated (gold).

Freshness

Last updated 2026-06-12 08:01:32; freshness should be verified.

Geography

Likely focuses on Tajikistan or Tajik-speaking regions, but specific geography is not stated.