---
license: mit
task_categories:
  - text-to-speech
  - automatic-speech-recognition
language:
  - ur
tags:
  - urdu
  - phonemes
  - kanade
  - speech
  - ipa
  - rasa
size_categories:
  - 1K<n<10K
---

# Rasa-Annotated

Enhanced version of [Rasa](https://huggingface.co/datasets/humair025/Rasa) with phoneme annotations and Kanade tokenizer features.

## Dataset Statistics

- **Total Samples**: 26,102
- **Total Duration**: 46.92 hours
- **Average Duration**: 6.47 seconds
- **Duration Range**: 0.31s - 45.34s
- **Average Phonemes**: 18.4 per sample
- **Average Kanade Tokens**: 530.7 per sample
- **Global Embedding Dimension**: 128

### Gender Distribution

| Gender | Count |
|--------|-------|
| Female | 12,583 |
| Male | 13,519 |

### Style Distribution

| Style | Count |
|-------|-------|
| CONV | 5,013 |
| FEAR | 765 |
| WIKI | 6,606 |
| BOOK | 2,040 |
| NEWS | 1,897 |
| DISGUST | 868 |
| ANGER | 814 |
| UMANG | 188 |
| PROPER NOUN | 4,595 |
| SAD | 736 |
| HAPPY | 839 |
| ALEXA | 470 |
| BB | 259 |
| SURPRISE | 824 |
| DIGI | 188 |

## New Columns

This dataset adds the following annotation columns (audio and wav_path are **NOT** included to save space):

1. **phonemes** (string): IPA phonemes with stress markers (space-separated)
2. **phonemes_no_stress** (string): IPA phonemes without stress markers
3. **phonemes_list** (list): List of individual phonemes with stress
4. **phoneme_count** (int): Number of phonemes
5. **content_token_indices** (list): Kanade content token indices
6. **global_embedding** (list): Kanade global embedding vector
7. **token_len** (int): Length of content token sequence

**Note**: The `audio` and `wav_path` columns are removed to reduce dataset size. Use the original [Rasa dataset](https://huggingface.co/datasets/humair025/Rasa) if you need the audio files.

## Generated By

- **Phonemes**: [urdu-g2p](https://github.com/zeeraktalat/urdu_g2p)
- **Kanade Features**: [kanade-12.5hz](https://huggingface.co/frothywater/kanade-12.5hz)
- **Processing**: CUDA-accelerated on GPU
- **Processing Date**: 2026-02-05

## Usage
```python
from datasets import load_dataset

dataset = load_dataset("humair025/Rasa-Annotated-25kHz")

# Example sample
sample = dataset['train'][0]
print(f"Text: {sample['text']}")
print(f"Duration: {sample['duration']}s")
print(f"Gender: {sample['gender']}")
print(f"Style: {sample['style']}")
print(f"Phonemes: {sample['phonemes']}")
print(f"Token length: {sample['token_len']}")
print(f"Global embedding shape: {len(sample['global_embedding'])}")

# Note: Audio is not included in this dataset
# Use the original Rasa dataset if you need audio files
```

## Original Dataset

Based on [Rasa](https://huggingface.co/datasets/humair025/Rasa)

## License

MIT License