dramabox-voice-acting-data-annotated

# DramaBox Voice Acting Data - Annotated Post-processed version of [laion/dramabox-voice-acting-data](https://huggingface.co/datasets/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](https://huggingface.co/nvidia/RE-USE) - Speech enhancement - [YatharthS/LavaSR](https://github.com/ysharma3501/LavaSR) - Super resolution - [openai/whisper](https://github.com/openai/whisper) turbo - ASR - [laion/voiceclap-large](https://huggingface.co/laion/voiceclap-large) - Voice embeddings - [ResembleAI/DramaBox](https://huggingface.co/ResembleAI/Dramabox) - Source TTS model