laion/dramabox-voice-acting-data-annotated

DramaBox Voice Acting Data - Annotated

Post-processed version of laion/dramabox-voice-acting-data.

Processing Pipeline

1. RE-USE Speech Enhancement (nvidia/RE-USE) - Applied to non-singing samples for noise reduction
2. LavaSR Super Resolution (YatharthS/LavaSR) - Audio bandwidth extension to 48kHz
3. Whisper Turbo ASR - Full transcript with word-level timestamps
4. Scene Split - Audio split at CUT TO: transition into two parts (Part 1 + Part 2)
5. VoiceCLAP Large Embeddings (laion/voiceclap-large) - 3584-dim voice embeddings per part

Purpose

Each DramaBox sample contains two emotional scenes from the same speaker separated by "CUT TO:". By splitting into parts, we create paired clips of the same voice identity with different emotions, enabling training of voice acting systems that maintain speaker consistency across emotional transitions.

Format

• MP3 files: 256kbps mono 48kHz
• Annotation JSON per sample with ASR, timestamps, embeddings, and metadata
• Packed in tar files matching the source dataset structure

Files per sample

• {prompt_id}_seed{NN}_part1.mp3 - Scene 1 audio
• {prompt_id}_seed{NN}_part2.mp3 - Scene 2 audio
• {prompt_id}_seed{NN}.json - Full annotations

Models Used

• nvidia/RE-USE - Speech enhancement
• YatharthS/LavaSR - Super resolution
• openai/whisper turbo - ASR
• laion/voiceclap-large - Voice embeddings
• ResembleAI/DramaBox - Source TTS model