TTS-AGI/voice-taxonomy-pretrain

Voice Taxonomy Pre-training Dataset

318,729 speech samples annotated with 57 voice taxonomy dimensions (0-6 ordinal scale) by a Whisper ensemble (4 models voting). Designed as pre-training data for voice attribute classifiers.

Related Datasets

Dataset	Purpose	Link
This dataset	Pre-training (large, noisy labels)	—
Fine-tuning (balanced, Gemini Flash)	Fine-tuning	TTS-AGI/voice-taxonomy-flash-train
Validation (Gemini 3.1 Pro gold)	Evaluation	TTS-AGI/voice-taxonomy-val

Format

WebDataset TAR with MP3+JSON pairs:

{stem}.mp3   # Audio (mono, 44.1kHz, 64kbps, ≤30s)
{stem}.json  # 57-dim taxonomy annotation

Each JSON:

{
  "AGEV": {"value": 3, "name": "Perceived Age", "label": "young adult"},
  "GEND": {"value": 5, "name": "Gender Presentation", "label": "standard masculine"},
  ...
}

Training Plan

See TRAINING_PLAN.md for the full training strategy (pre-train → fine-tune → evaluate) and train_voice_taxonomy.py for a self-contained training script.

Quick Start

# Download
huggingface-cli download TTS-AGI/voice-taxonomy-pretrain --local-dir .

# Pre-train
python train_voice_taxonomy.py --phase pretrain --encoder laion/BUD-E-Whisper --gpu 0

Taxonomy

57 dimensions covering: speaker identity, timbral quality, resonance placement, prosody, articulation, emotion, and speaking style. Each rated 0-6. See taxonomy_labels.json for full definitions.

Labels

Labels were generated by a Whisper ensemble (4 BUD-E-Whisper variants voting). These are noisier than the Gemini-annotated fine-tuning and validation sets, but the 10x larger dataset size compensates.