Datasets:

Modalities:
Audio
Text
Formats:
parquet
Languages:
English
ArXiv:
Tags:
audio
Libraries:
Datasets
Dask
License:
BorisovMaksim commited on
Commit
c457ed6
Β·
verified Β·
1 Parent(s): 7fa3226

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +114 -0
README.md CHANGED
@@ -0,0 +1,114 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # NonverbalTTS Dataset πŸŽ΅πŸ—£οΈ
2
+
3
+ [![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.15274617.svg)](https://doi.org/10.5281/zenodo.15274617)
4
+ [![Hugging Face](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Dataset-blue)](https://huggingface.co/datasets/deepvk/NonverbalTTS)
5
+
6
+ **NonverbalTTS** is a 17-hour open-access English speech corpus with aligned text annotations for **nonverbal vocalizations (NVs)** and **emotional categories**, designed to advance expressive text-to-speech (TTS) research.
7
+
8
+ ## Key Features ✨
9
+
10
+ - **17 hours** of high-quality speech data
11
+ - **10 NV types**: Breathing, laughter, sighing, sneezing, coughing, throat clearing, groaning, grunting, snoring, sniffing
12
+ - **8 emotion categories**: Angry, disgusted, fearful, happy, neutral, sad, surprised, other
13
+ - **Diverse speakers**: 2296 speakers (60% male, 40% female)
14
+ - **Multi-source**: Derived from [VoxCeleb](https://www.robots.ox.ac.uk/~vgg/data/voxceleb/) and [Expresso](https://arxiv.org/abs/2308.05725) corpora
15
+ - **Rich metadata**: Emotion labels, NV annotations, speaker IDs, audio quality metrics
16
+
17
+ <!-- ## Dataset Structure πŸ“‚
18
+
19
+ NonverbalTTS/
20
+ β”œβ”€β”€ wavs/ # Audio files (16-48kHz WAV format)
21
+ β”‚ β”œβ”€β”€ ex01_sad_00265.wav
22
+ β”‚ └── ...
23
+ β”œβ”€β”€ .gitattributes
24
+ β”œβ”€β”€ README.md
25
+ └── metadata.csv # Metadata annotations -->
26
+
27
+
28
+ ## Metadata Schema (`metadata.csv`) πŸ“‹
29
+
30
+ | Column | Description | Example |
31
+ |--------|-------------|---------|
32
+ | `index` | Unique sample ID | `ex01_sad_00265` |
33
+ | `file_name` | Audio file path | `wavs/ex01_sad_00265.wav` |
34
+ | `Emotion` | Emotion label | `sad` |
35
+ | `Initial text` | Raw transcription | `"So, Mom, 🌬️ how've you been?"` |
36
+ | `Annotator response {1,2,3}` | Refined transcriptions | `"So, Mom, how've you been?"` |
37
+ | `Result` | Final fused transcription | `"So, Mom, 🌬️ how've you been?"` |
38
+ | `dnsmos` | Audio quality score (1-5) | `3.936982` |
39
+ | `duration` | Audio length (seconds) | `3.6338125` |
40
+ | `speaker_id` | Speaker identifier | `ex01` |
41
+ | `data_name` | Source corpus | `Expresso` |
42
+ | `gender` | Speaker gender | `m` |
43
+
44
+ **NV Symbols**: 🌬️=Breath, πŸ˜‚=Laughter, etc. (See [Annotation Guidelines](https://zenodo.org/records/15274617))
45
+
46
+ ## Loading the Dataset πŸ’»
47
+
48
+ ```python
49
+ from datasets import load_dataset
50
+
51
+ dataset = load_dataset("deepvk/NonverbalTTS")
52
+ ```
53
+
54
+ <!-- # Access train split
55
+ ```print(dataset["train"][0])```
56
+
57
+ # Output: {'index': 'ex01_sad_00265', 'file_name': 'wavs/ex01_sad_00265.wav', ...}
58
+ -->
59
+
60
+ ## Annotation Pipeline πŸ”§
61
+
62
+ 1. **Automatic Detection**
63
+ - NV detection using [BEATs](https://arxiv.org/abs/2409.09546)
64
+ - Emotion classification with [emotion2vec+](https://arxiv.org/abs/2402.XXX)
65
+ - ASR transcription via Canary model
66
+
67
+ 2. **Human Validation**
68
+ - 3 annotators per sample
69
+ - Filtered non-English/multi-speaker clips
70
+ - NV/emotion validation and refinement
71
+
72
+ 3. **Fusion Algorithm**
73
+ - Majority voting for final transcriptions
74
+ - Pyalign-based sequence alignment
75
+ - Multi-annotator hypothesis merging
76
+
77
+
78
+ ## Benchmark Results πŸ“Š
79
+
80
+
81
+ Fine-tuning CosyVoice-300M on NonverbalTTS achieves parity with state-of-the-art proprietary systems:
82
+ |Metric | NVTTS | CosyVoice2 |
83
+ | ------- | ------- | ------- |
84
+ |Speaker Similarity | 0.89 | 0.85 |
85
+ |NV Jaccard (Laugh) | 0.92 | 0.74 |
86
+ |Human Preference | 33.4% | 35.4% |
87
+
88
+
89
+ ## Use Cases πŸ’‘
90
+ - Training expressive TTS models
91
+ - Zero-shot NV synthesis
92
+ - Emotion-aware speech generation
93
+ - Prosody modeling research
94
+
95
+ ## License πŸ“œ
96
+ - Annotations: CC BY-NC-SA 4.0
97
+ - Audio: Adheres to original source licenses (VoxCeleb, Expresso)
98
+
99
+
100
+ ## Citation πŸ“
101
+
102
+ ```
103
+ @dataset{nonverbaltts2024,
104
+ author = {Anonymous},
105
+ title = {NonverbalTTS Dataset},
106
+ month = December,
107
+ year = 2024,
108
+ publisher = {Zenodo},
109
+ version = {1.0},
110
+ doi = {10.5281/zenodo.15274617},
111
+ url = {https://zenodo.org/records/15274617}
112
+ }
113
+ ```
114
+