ChristophSchuhmann's picture
Upload README.md with huggingface_hub
f5d809b verified
|
Raw History Blame Contribute Delete
2.76 kB
metadata
license: cc-by-4.0
task_categories:
  - audio-classification
  - text-to-speech
tags:
  - voice
  - speaker-embedding
  - deduplicated
  - emotion
  - quality-filtered
pretty_name: AI Voices Deduplicated
size_categories:
  - 1K<n<10K

AI Voices Deduplicated

2,004 high-quality, speaker-deduplicated AI voice samples from laion/reference_ai_voices_with_timbre_annotations.

Dataset Summary

  • Source: 17 tar files (~32,000 samples) from laion/reference_ai_voices_with_timbre_annotations
  • Quality filtering: score_background_quality >= 3.5 AND score_content_enjoyment >= 5.0 (11,473 passed)
  • Speaker deduplication: Speaker embeddings (Orange/Speaker-wavLM-tbr) + agglomerative clustering (cosine distance threshold 0.0999), keeping highest quality sample per cluster
  • Final count: 2,004 unique voice samples

Structure

ai-voices-deduplicated.tar
├── male/
│   ├── 02_child/       (1 sample)
│   ├── 03_teenager/    (5 samples)
│   ├── 04_young_adult/ (217 samples)
│   ├── 05_adult/       (813 samples)
│   └── 08_elderly/     (1 sample)
├── female/
│   ├── 02_child/       (16 samples)
│   ├── 03_teenager/    (3 samples)
│   ├── 04_young_adult/ (376 samples)
│   ├── 05_adult/       (514 samples)
│   └── 08_elderly/     (1 sample)
└── androgynous/
    ├── 02_child/       (13 samples)
    ├── 04_young_adult/ (19 samples)
    └── 05_adult/       (25 samples)

Each audio file has a corresponding .json sidecar with:

Gender Distribution

Gender Count
Male 1,037
Female 910
Androgynous 57

Quality Scores

All samples have:

  • score_background_quality >= 3.5 (DNS MOS background quality)
  • score_content_enjoyment >= 5.0 (content enjoyment rating)

Processing Pipeline

  1. Downloaded all 17 tar files from source dataset
  2. Annotated with Empathic Insight Voice Plus (55 emotion + 4 quality scores)
  3. Filtered by quality thresholds
  4. Generated speaker embeddings using Orange/Speaker-wavLM-tbr (timbre-based representation)
  5. Agglomerative clustering per gender group with cosine distance
  6. Binary search for threshold to target ~2,000 samples
  7. Kept highest quality sample per speaker cluster