README.md

---
language:
- ml
license: cc-by-sa-4.0
task_categories:
- automatic-speech-recognition
- text-to-speech
tags:
- malayalam
- speech
- audio
pretty_name: Malayalam Speech Dataset (VibeVoice)
size_categories:
- 1K<n<10K
---

# Malayalam Speech Dataset (VibeVoice)

This dataset contains Malayalam speech audio files and their corresponding transcriptions, prepared for fine-tuning ASR (Automatic Speech Recognition) models like VibeVoice.

## Dataset Description

The dataset consists of approximately 4,126 Malayalam audio clips, split into training and testing sets with a 90/10 ratio, stratified by speaker gender.

- **Total Audio Files**: 4,126
- **Train Split**: 3,712 samples
- **Test Split**: 414 samples
- **Total Duration**: ~5-6 hours (estimated)
- **Format**: 16kHz WAV
- **Annotation**: JSON segments and split-specific `metadata.jsonl` entries.

## Dataset Structure

The dataset is organized into `train` and `test` directories:
1. `train/`: Contains 3,712 audio-JSON pairs and `metadata.jsonl`.
2. `test/`: Contains 414 audio-JSON pairs and `metadata.jsonl`.

Each split contains:
- `[id].wav`: The audio file.
- `[id].json`: Segment information and transcription.
- `metadata.jsonl`: Mapping of audio files to transcriptions.

### VibeVoice JSON format
Each audio file is accompanied by a JSON file with the following structure:
```json
{
  "audio_duration": 2.523,
  "audio_path": "female_mlf_01130_00015565294.wav",
  "segments": [
    {
      "start": 0.0,
      "end": 2.523,
      "text": "അങ്ങനെ രണ്ടാം ലോകമഹായുദ്ധം അവസാനിച്ചു",
      "speaker": 0
    }
  ]
}
```

### metadata.jsonl format
```json
{"file_name": "female_mlf_01130_00015565294.wav", "transcription": "അങ്ങനെ രണ്ടാം ലോകമഹായുദ്ധം അവസാനിച്ചു"}
```

## How to use

```python
from datasets import load_dataset

dataset = load_dataset("Arjunj/malayalam_speech", split="train")

# Access a sample
sample = dataset[0]
print(sample["audio"])
print(sample["transcription"])
```

## Source
This dataset was processed and collected from various Malayalam TTS/ASR sources, including OpenSLR IDs 63 and 64, formatted specifically for VibeVoice fine-tuning requirements.

## License
Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0).