raditotev/bg-audiobooks-tts

Bulgarian Audiobook Speech Dataset

A high-quality Bulgarian speech dataset derived from audiobooks narrated by Plamen Sivov, suitable for text-to-speech (TTS) and automatic speech recognition (ASR) tasks.

Dataset Summary

Property	Value
Language	Bulgarian (bg)
Total Duration	15.2 hours
Total Clips	10,627
Speaker	Plamen Sivov (single speaker)
Source	YouTube audiobooks
Sample Rate	24,000 Hz
Audio Format	WAV, mono, 16-bit PCM
Clip Duration	3–15 seconds
License	CC0-1.0

Dataset Structure

Each sample contains:

Field	Type	Description
file_name	string	Path to audio file
text	string	Raw transcription (Whisper large-v3)
text_normalized	string	TTS-ready text (numbers/dates/abbr expanded)
duration	float	Audio duration in seconds
speaker_id	string	Speaker identifier
speaker_name	string	Speaker full name
language	string	Language code (bg)
source_file	string	Original audiobook file
book_title	string	Book/chapter title
segment_start	float	Start time in source file (seconds)
segment_end	float	End time in source file (seconds)
rms_energy	float	Audio energy (loudness indicator)
sample_rate	int	Audio sample rate (24000)
clip_id	int	Unique clip identifier

Usage

from datasets import load_dataset

# Load dataset
ds = load_dataset("raditotev/bg-audiobooks-tts")

# Access a sample
sample = ds["train"][0]
print(sample["text"])
print(sample["text_normalized"])
print(f"Duration: {sample['duration']:.1f}s")

# Play audio (in notebook)
from IPython.display import Audio
Audio(sample["file_name"], rate=24000)

Use Cases

• Text-to-Speech (TTS): Fine-tune models like F5-TTS, XTTS-v2, or VITS for Bulgarian
• Automatic Speech Recognition (ASR): Fine-tune Whisper or other ASR models for Bulgarian
• Voice Cloning: Single-speaker dataset ideal for voice adaptation
• Bulgarian NLP: Text normalization research, prosody analysis

Dataset Creation

Source

Audio from Plamen Sivov's YouTube channel, used with explicit permission from the author.

Processing Pipeline

1. Audio downloaded with yt-dlp
2. Converted to 24kHz mono WAV
3. Transcribed with OpenAI Whisper large-v3 (MLX)
4. Segmented into 3–15 second clips
5. Filtered for quality (minimum energy, text length, Cyrillic ratio)
6. Text normalized for TTS (numbers, dates, abbreviations expanded to words)

Quality Notes

• Transcriptions generated by Whisper large-v3 — expect ~95%+ accuracy for Bulgarian
• Some clips may contain minor transcription errors
• text_normalized field provides TTS-ready text with all non-verbal elements expanded

Citation

@misc{bg_audiobooks_tts_2026,
  title={Bulgarian Audiobook Speech Dataset},
  author={Radi Totev/ Plamen Sivov},
  year={2026},
  publisher={HuggingFace},
  howpublished={\url{https://huggingface.co/datasets/raditotev/bg-audiobooks-tts}}
}

Acknowledgments

• Plamen Sivov for creating and sharing the audiobook recordings
• OpenAI Whisper for transcription
• Apple MLX for efficient on-device processing

License

This dataset is released under CC0-1.0. Attribution should be given to Plamen Sivov as the original narrator.