---
language:
  - as
  - bn
  - en
  - gu
  - hi
  - kn
  - ml
  - mr
  - ne
  - or
  - pa
  - ta
  - te
license: cc-by-4.0
task_categories:
  - text-to-speech
  - automatic-speech-recognition
size_categories:
  - 100K<n<1M
tags:
  - indic
  - multilingual
  - tts
  - speech
---

# Processed TTS Multilingual Data

Validated and quality-checked multilingual speech datasets for TTS training, covering 12+ Indian languages.

## Datasets Included

| Subset | Samples | Hours | Description |
|---|---|---|---|
| `indic_voices_r` | 239,684 | 548.8h | Indic Voices_R — IVR recordings |
| `rasa` | 201,509 | 361.2h | RASA — read speech (wiki, conv, book, news) |
| `indictts_iitm` | 155,236 | 253.6h | Indic TTS (IIT Madras) — studio TTS recordings at 48kHz |
| **Total** | **596,429** | **1,163.6h** | |

## Languages

Assamese (as), Bengali (bn), English (en), Gujarati (gu), Hindi (hi), Kannada (kn), Malayalam (ml), Marathi (mr), Nepali (ne), Odia (or), Punjabi (pa), Tamil (ta), Telugu (te)

## Structure

```
├── indic_voices_r/
│   ├── metadata.csv
│   └── audio/{lang}/*.wav
├── rasa/
│   ├── metadata.csv
│   └── audio/{lang}/*.wav
└── indictts_iitm/
    ├── metadata.csv
    └── audio/{lang}/*.wav
```

## Schema (metadata.csv)

Each subset has a `metadata.csv` with these columns:

| Field | Description |
|---|---|
| `file_name` | Relative path to audio file (e.g., `audio/bn/file.wav`) |
| `text` | Transcript text |
| `lang` | ISO 639-1 language code |
| `speaker_id` | Speaker identifier |
| `duration` | Audio duration in seconds |
| `source` | Original data source |
| `emotion` | Emotion label |
| `domain` | Text domain (wiki, conv, book, news, etc.) |
| `snr_db` | Signal-to-noise ratio in dB |
| `silence_ratio` | Fraction of silent frames |
| `clipping_ratio` | Fraction of clipped samples |

## Quality Checks Applied

All data has been validated through a 4-check pipeline:
1. **SNR + Silence + Duration** — reject low SNR (<10dB), excess silence (>35%), out-of-range duration (<1.5s or >30s), clipping (>1%)
2. **Speaking Rate** — reject abnormal speaking rates (<2 or >25 chars/sec)
3. **Text Normalization** — Unicode NFC normalization applied
4. **Audio Corruption** — reject empty, all-zeros, NaN/Inf, DC offset >0.1

## Usage

```python
from datasets import load_dataset

# Load a specific subset
ds = load_dataset(
    "PalakEngineerMaster/Processed_TTS_Multilingual_Data",
    data_dir="rasa",
    split="train",
)

# Access a sample
sample = ds[0]
print(sample["text"])
# audio is at sample["file_name"]
```

## Audio Format

- Format: WAV
- Sample rate: 16kHz (Indic Voices_R, RASA) / 48kHz (Indic TTS IIT M)
- Channels: mono