Hindi HQ: 48,000 Audio-Text Pairs for Speech Synthesis

48,078 audio-text pairs totaling 89.63 hours of Hindi speech, collected from YouTube using auto-generated captions for transcription. The dataset is heavily skewed, with one speaker contributing 76.9% of the content. Created by user somu9 and last updated on June 8, 2026.

Use Cases

• Train Hindi text-to-speech models based on the audio-text pairs.
• Fine-tune Hindi speech recognition systems based on the transcribed audio segments.
• Analyze speaker characteristics and prosody based on the multi-speaker distribution.
• Benchmark speech synthesis quality using the 24kHz mono audio format.

Strengths

• 48,078 rows provide a substantial volume of training data.
• 89.63 total hours of audio offers a significant duration for model training.
• Audio is stored in a modern, efficient format as sharded parquet files with embedded audio.
• Speaker distribution statistics are provided, showing contributions from at least 312 distinct video sources.

Limitations

• Speaker distribution is highly imbalanced, with one speaker (spk0000) accounting for 76.9% of the data, which may introduce bias.
• Transcriptions rely on YouTube's auto-generated closed captions, which may contain errors.
• Column-level documentation and sample data are unavailable, making field semantics difficult to assess prior to download.

Provenance

Source

YouTube videos.

Collection Method

Collected from YouTube using auto-generated closed captions for transcription.

Freshness

Last updated 2026-06-08 20:36:42