ahk-d/peaky-blinders-learning-purpose-only

TTS Dataset - Peaky Blinders

This dataset contains audio segments with transcriptions from Peaky Blinders for Text-to-Speech training.

Dataset Generation

This dataset was generated using the TTS-Dataset-Maker pipeline, which provides:

• Silero VAD-based silence removal - Removes long silences while preserving natural speech gaps
• DeepFilterNet denoising - CPU-optimized audio denoising with gentle attenuation (15dB)
• AssemblyAI transcription - High-quality speech-to-text with speaker diarization
• Existing transcript support - Uses pre-existing transcripts to skip transcription
• Precise segmentation - Exact audio segment extraction matching transcript timings
• Unique file naming - Source-aware segment names to prevent collisions

Processing Pipeline

1. Audio preprocessing with DeepFilterNet denoising (atten_lim_db=15.0 for gentle noise reduction)
2. Voice Activity Detection using Silero VAD to remove long silences
3. Speaker diarization and transcription (or loading existing transcripts)
4. Precise audio segmentation based on transcript timestamps
5. Quality validation and filtering of segments

Dataset Structure

• audio/: Audio files (WAV format, 44.1kHz)
• metadata.json: Complete segment metadata for Label Studio
• transcripts/: Per-file transcript JSON files

Usage

import json
import os

# Load metadata
with open("metadata.json", 'r') as f:
    data = json.load(f)

# Access segments
for segment in data:
    print(f"Text: {segment['text']}")
    print(f"Speaker: {segment['speaker_id']}")
    print(f"Audio: {segment['audio_file']}")
    print(f"Duration: {segment['duration']}s")
    print(f"Quality Score: {segment['quality_score']}")

Dataset Details

• Language: English
• Format: WAV audio files (44.1kHz, 16-bit PCM)
• Speakers: Multiple speakers (A, B, C, D, E)
• Duration: Variable segment lengths (filtered for >1s duration)
• Quality: High-quality audio with quality scores and confidence metrics
• Total Segments: 356 segments
• Total Duration: ~89 minutes
• Processing: Denoised with DeepFilterNet, silence-removed with Silero VAD

Metadata Format

Each segment contains:

• text: Transcribed text
• speaker_id: Speaker identifier
• audio_file: Path to audio file
• duration: Duration in seconds
• quality_score: Audio quality score
• confidence_score: Transcription confidence
• start_time: Start time in milliseconds
• end_time: End time in milliseconds
• source_file: Original source file path

License

MIT License

Citation

If you use this dataset, please cite: @dataset{peaky_blinders_tts, title={TTS Dataset - Peaky Blinders}, author={ahk-d}, year={2024}, url={https://huggingface.co/datasets/ahk-d/peaky-blinders-learning-purpose-only} }

Generation Tool

Generated using TTS-Dataset-Maker - A comprehensive pipeline for creating high-quality TTS datasets from audio files with automatic transcription, denoising, and segmentation.