---
dataset_info:
  features:
  - name: audio
    dtype: audio
  - name: sentence
    dtype: string
  - name: phonemes
    dtype: string
  splits:
  - name: train
    num_bytes: 754910024.376
    num_examples: 3336
  - name: test
    num_bytes: 185837173
    num_examples: 834
  download_size: 885325061
  dataset_size: 940747197.376
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
license: mit
task_categories:
- text-to-speech
- automatic-speech-recognition
language:
- dv
tags:
- dhivehi
- divehi
- speech
- voice
pretty_name: kn_audio_dataset
size_categories:
- 1K<n<10K
---
# Dhivehi Audio Dataset

This is a Dhivehi (Maldivian) speech synthesis dataset with audio recordings, text transcriptions, and phonetic annotations. All recordings are by one speaker.

![Dataset Analysis](assets/analysis.png)

## Dataset Overview

This dataset provides 4,170 high-quality audio samples in Dhivehi.

### Key Statistics

| Metric | Value |
|--------|-------|
| **Total Audio Files** | 4,170 |
| **Total Duration** | 5.87 hours (352.0 minutes) |
| **Average Clip Length** | 5.06 seconds |
| **Total Words** | 30,068 |
| **Unique Phonemes** | 14190 |
| **Unique Characters** | 62 |

### Audio Specifications

- **Sample Rate**: 22050 Hz
- **Channels**: 1 (Mono/Stereo)
- **Bit Depth**: 16 bit
- **Format**: WAV (uncompressed)


## Data Distribution

### Duration Analysis

- **Shortest clip**: 1.18 seconds
- **Longest clip**: 19.08 seconds  
- **Standard deviation**: 2.41 seconds

### Text Characteristics

- **Average sentence length**: 65.7 characters
- **Average words per sentence**: 7.2
- **Average phonemes per sentence**: 7.2
- **Sentence length range**: 6 - 198 characters

### Most Common Phonemes

| Phoneme | Count | Frequency |
|---------|-------|----------|
| `ʋes̺` | 221 | 0.73% |
| `e` | 212 | 0.70% |
| `eːn̪aː` | 205 | 0.68% |
| `ahaɾeŋ` | 170 | 0.56% |
| `ad̪i` | 160 | 0.53% |
| `ʋaɾaʔ` | 155 | 0.52% |
| `mi` | 149 | 0.50% |
| `eiː` | 146 | 0.49% |
| `n̪amaʋes̺` | 144 | 0.48% |
| `n̪uːŋ` | 140 | 0.47% |


## Common Voice 17.0 Comparison

This dataset was evaluated against the Dhivehi portion of Mozilla’s Common Voice 17.0 to assess overlap and uniqueness. The results are as follows:


| Metric | Value |
|--------|-------|
| **Common Voice Sentences** | 7,117 |
| **This Dataset Sentences** | 4,170 |
| **Overlapping Sentences** | 3,443 (82.6%) |
| **Unique to This Dataset** | 727 (17.4%) |

Although a substantial number of sentences are shared between the two datasets, each sentence in this dataset is recorded by a different speaker. This contributes to greater speaker diversity, making the dataset particularly suitable for training and evaluating speaker-independent models.

### Key Insights

* **17.4% Unique Sentences**: A notable portion of this dataset comprises sentences not present in Common Voice 17.0, contributing new linguistic content to the Dhivehi speech domain.
* **Speaker and Quality Variation**: Shared sentences offer a basis for evaluating speaker diversity and audio quality differences across datasets.
* **Complementary Value**: Serves as a valuable supplement to Common Voice 17.0 by providing additional recordings and speaker variability, enhancing the robustness of Dhivehi speech model training.

### Example Sentence from Common Voice 17.0 Compared to This Dataset

| **Common Voice 17.0**| **This Dataset**|
|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| <audio controls > <source src="https://huggingface.co/datasets/alakxender/dhivehi-audio-kn/resolve/main/assets/audio_samples/CV_000df3f32ae6beb89a968f52661c754a2080f8e785818bf319a22a15a1d9a853_validation.wav" type="audio/mpeg"> Your browser does not support the audio element. </audio> | <audio controls > <source src="https://huggingface.co/datasets/alakxender/dhivehi-audio-kn/resolve/main/assets/audio_samples/THIS_000df3f32ae6beb89a968f52661c754a2080f8e785818bf319a22a15a1d9a853.wav" type="audio/mpeg"> Your browser does not support the audio element. </audio> |
| <audio controls > <source src="https://huggingface.co/datasets/alakxender/dhivehi-audio-kn/resolve/main/assets/audio_samples/CV_007f4e661aa97e0b6d7ee579ffb1553926ce30586ff1779c801ab70c55db9683_test.wav" type="audio/mpeg"> Your browser does not support the audio element. </audio> | <audio controls > <source src="https://huggingface.co/datasets/alakxender/dhivehi-audio-kn/resolve/main/assets/audio_samples/THIS_007f4e661aa97e0b6d7ee579ffb1553926ce30586ff1779c801ab70c55db9683.wav" type="audio/mpeg"> Your browser does not support the audio element. </audio> |
| <audio controls > <source src="https://huggingface.co/datasets/alakxender/dhivehi-audio-kn/resolve/main/assets/audio_samples/CV_00324a363d8eb184dc279e576b373a85c70f0a23b3b26ea4307dc6091ae78898_validation.wav" type="audio/mpeg"> Your browser does not support the audio element. </audio> | <audio controls > <source src="https://huggingface.co/datasets/alakxender/dhivehi-audio-kn/resolve/main/assets/audio_samples/THIS_00324a363d8eb184dc279e576b373a85c70f0a23b3b26ea4307dc6091ae78898.wav" type="audio/mpeg"> Your browser does not support the audio element. </audio> |

## Use Cases

This dataset is suitable for:

- **Text-to-Speech (TTS) synthesis** - Training neural speech synthesis models
- **Automatic Speech Recognition (ASR)** - Dhivehi speech recognition systems
- **Phonetic research** - Analysis of Dhivehi phonology and pronunciation
- **Language preservation** - Digital archiving of Dhivehi speech patterns
- **Cross-lingual studies** - Comparative phonetic analysis

## Loading the Dataset

### Using 🤗 Datasets

```python
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("alakxender/dhivehi-audio-kn")

# Access training split
train_data = dataset["train"]

# Example: Get first sample
sample = train_data[0]
print(f"Text: {sample['sentence']}")
print(f"Phonemes: {sample['phonemes']}")
# Audio is available in sample['audio']
```

## Acknowledgments

Special thanks to [@kudanai](https://huggingface.co/kudanai) for providing the dataset used as the source. This dataset retains the same speaker, offering consistent voice characteristics across all recordings.