Hugging Face's logo

Datasets:
nate0001
/
KionTTS

Tasks:
Text-to-Speech
Modalities:
Audio
Text
Formats:
parquet
Languages:
English
Size:
10K - 100K
Tags:
audio
tts
Synthetic
emotion
kion
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
nate0001 commited on
Commit
069362b
·
verified ·
1 Parent(s): affe489

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +18 -23
README.md CHANGED
@@ -27,13 +27,6 @@ dataset_info:
27
  dtype: string
28
  - name: styles
29
  dtype: string
30
- splits:
31
- - name: tier1
32
- num_examples: 3323
33
- - name: tier2
34
- num_examples: 933
35
- - name: tier3
36
- num_examples: 14517
37
  configs:
38
  - config_name: default
39
  data_files:
@@ -52,27 +45,29 @@ generated with IndexTTS2 using emotion and style conditioning.
52
 
53
  ## Dataset Structure
54
 
55
- Each row contains a spoken utterance with its text transcript, emotion/style tags, and audio.
56
-
57
  | Field | Type | Description |
58
  |-------|------|-------------|
59
- | uid=1000(nate) gid=1000(nate) groups=1000(nate),950(docker),954(libvirt),984(uucp),990(kvm),998(wheel),1001(sdkusers),1002(Kion),1003(caelestia),1004(Ki0n),1005(dialout) | string | Unique 8-char hex ID |
60
- | | string | Quality tier: (regenerated), (spot-check), (clean) |
61
- | | string | Full text with inline emotion tags e.g. |
62
- | | Audio | WAV audio, 24kHz mono |
63
- | | string | JSON dict of emotion labels and intensities |
64
- | | string | JSON dict of style labels and intensities |
65
 
66
  ## Tiers
67
 
68
- - **Tier 1** (3,323 files) — Regenerated with fixed to eliminate reference speaker leakage
69
- - **Tier 2** (933 files) — Medium-risk files, recommended for spot-checking before use
70
- - **Tier 3** (14,517 files) — Clean files, safe for direct training use
 
 
71
 
72
- ## Emotion / Style Tags
 
73
 
74
- Emotions: , , , , , , ,
75
- , , , , , ,
76
 
77
- Styles: , , , , , ,
78
- , , ,
 
 
27
  dtype: string
28
  - name: styles
29
  dtype: string
 
 
 
 
 
 
 
30
  configs:
31
  - config_name: default
32
  data_files:
 
45
 
46
  ## Dataset Structure
47
 
 
 
48
  | Field | Type | Description |
49
  |-------|------|-------------|
50
+ | `id` | string | Unique 8-char hex ID |
51
+ | `tier` | string | `tier1` (regenerated fixes), `tier2` (spot-check), `tier3` (clean) |
52
+ | `text` | string | Full text with inline emotion tags e.g. `[happy=0.7] Hello!` |
53
+ | `audio` | Audio | WAV audio, 24kHz mono |
54
+ | `emotions` | string | JSON dict of emotion labels and intensities |
55
+ | `styles` | string | JSON dict of style labels and intensities |
56
 
57
  ## Tiers
58
 
59
+ - **Tier 1** (3,323 files) — Regenerated with corrected emo_alpha to fix reference speaker leakage
60
+ - **Tier 2** (933 files) — Medium-risk; recommend spot-checking before training use
61
+ - **Tier 3** (14,517 files) — Clean; safe for direct training use
62
+
63
+ ## Loading the dataset
64
 
65
+ ```python
66
+ from datasets import load_dataset
67
 
68
+ # Load clean tier only
69
+ ds = load_dataset("nate0001/KionTTS", split="tier3")
70
 
71
+ # Load all tiers
72
+ ds = load_dataset("nate0001/KionTTS")
73
+ ```