---
task_categories:
- automatic-speech-recognition
language:
- ta
- gu
- te
dataset_info:
- config_name: gu
  features:
  - name: audio
    dtype: audio
  - name: transcription
    dtype: string
  - name: audio_id
    dtype: string
  splits:
  - name: train
    num_bytes: 4739983146.192
    num_examples: 22807
  - name: test
    num_bytes: 578740547.85
    num_examples: 3075
  download_size: 4892009740
  dataset_size: 5318723694.042001
- config_name: ta
  features:
  - name: audio
    dtype: audio
  - name: transcription
    dtype: string
  - name: audio_id
    dtype: string
  splits:
  - name: train
    num_bytes: 8359705604.568
    num_examples: 39131
  - name: test
    num_bytes: 603047036.672
    num_examples: 3081
  download_size: 5128373412
  dataset_size: 8962752641.24
- config_name: te
  features:
  - name: audio
    dtype: audio
  - name: transcription
    dtype: string
  - name: audio_id
    dtype: string
  splits:
  - name: train
    num_bytes: 8420579336.816
    num_examples: 44882
  - name: test
    num_bytes: 579072608.12
    num_examples: 3040
  download_size: 5096987441
  dataset_size: 8999651944.936
configs:
- config_name: gu
  data_files:
  - split: train
    path: gu/train-*
  - split: test
    path: gu/test-*
- config_name: ta
  data_files:
  - split: train
    path: ta/train-*
  - split: test
    path: ta/test-*
- config_name: te
  data_files:
  - split: train
    path: te/train-*
  - split: test
    path: te/test-*
---

# Microsoft Speech Corpus – Indian Languages

## Dataset Description

This dataset is a redistribution of the **Microsoft Speech Corpus (Indian Languages)** containing conversational and phrasal speech training and test data for **Telugu**, **Tamil**, and **Gujarati** languages. Each entry includes an audio recording and its corresponding transcript.

> **Attribution required:** *"Data provided by Microsoft and SpeechOcean.com"*

> ⚠️ **License:** This data is provided for **research purposes only**. Commercial use is strictly prohibited.

---

## Dataset Structure

### Configurations

| Config | Language | Script |
|--------|----------|--------|
| `ta`   | Tamil    | Tamil script (தமிழ்) |
| `gu`   | Gujarati | Gujarati script (ગુજરાતી) |
| `te`   | Telugu   | Telugu script (తెలుగు) |

### Splits & Size

| Language | Split | Examples |
|----------|-------|----------|
| Tamil (`ta`)    | train | 39,131 |
| Tamil (`ta`)    | test  | 3,081  |
| Gujarati (`gu`) | train | 22,807 |
| Gujarati (`gu`) | test  | 3,075  |
| Telugu (`te`)   | train | 44,882 |
| Telugu (`te`)   | test  | 3,040  |

**Total: 116,016 examples**

### Data Fields

| Field | Type | Description |
|-------|------|-------------|
| `audio` | `Audio` | Audio file with sampling rate metadata |
| `transcription` | `string` | Transcript of the spoken content |
| `audio_id` | `string` | Unique identifier of the audio file |

---

## Usage

```python
from datasets import load_dataset

# Load Tamil
ta = load_dataset("deepdml/microsoft-speech-corpus-indian", "ta", trust_remote_code=True)

# Load Gujarati
gu = load_dataset("deepdml/microsoft-speech-corpus-indian", "gu", trust_remote_code=True)

# Load Telugu
te = load_dataset("deepdml/microsoft-speech-corpus-indian", "te", trust_remote_code=True)
```

Access a sample:

```python
sample = ta["train"][0]
print(sample["transcription"])
# → "நீங்க அழிகா இருக்கீங்க என்று சொல்வது..."

sample["audio"]
# → {"array": [...], "sampling_rate": 16000, "path": "000020021.wav"}
```

---

## Dataset Creation

### Source Data

The original data was collected and released by **Microsoft Research** in collaboration with **SpeechOcean.com** as part of the Microsoft Speech Corpus for Indian Languages. The corpus covers conversational and phrasal speech scenarios.

### Collection Process

Audio recordings were collected from native speakers of Tamil, Gujarati, and Telugu across India and transcribed professionally. The data reflects natural spoken language including conversational speech patterns.

---

## License & Attribution

This dataset is made available for **non-commercial research use only**.

If you publish findings using this dataset, you **must** include the following attribution:

> *"Data provided by Microsoft and SpeechOcean.com"*

Original dataset: [Microsoft Research Open Data](https://www.microsoft.com/en-us/download/details.aspx?id=105292)

---

## Citation

```bibtex
@dataset{microsoft_speech_corpus_indian,
  title     = {Microsoft Speech Corpus (Indian Languages)},
  author    = {Microsoft Research and SpeechOcean.com},
  year      = {2020},
  url       = {https://msropendata.com/datasets/7230b4b1-912d-400e-be58-f84e0512985e},
  note      = {Non-commercial research use only}
}
```