---
license: apache-2.0
task_categories:
  - text-to-speech
language:
  - hi
  - en
tags:
  - hindi
  - hinglish
  - tts
  - soniox
  - kupe
pretty_name: kupe-tts
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/clusters/cluster-*.parquet
---

# kupe-tts

Self-contained Hindi / Hinglish TTS **parquet clusters**
(embedded audio bytes + text + full STT char timestamps).

## Default split (`train`)

`data/clusters/cluster-XXXXX-of-00010.parquet` —
**10** clusters, **30** rows.

| column | type | description |
|--------|------|-------------|
| `audio` | Audio | **embedded wav bytes** (playable in Viewer) |
| `text` | string | utterance transcript |
| `index` | int64 | corpus index |
| `meta_json` | string | **full** Soniox generation + timestamps JSON |
| `characters_json` | string | char-level STT timestamps |
| `word_timestamps_json` | string | word-level timestamps |
| `duration_sec` | float | audio duration |
| `sample_rate` | int64 | Hz (24000) |
| `language` / `voice` / `model` | string | generation params |
| `coverage_pct` | float | timestamp coverage |

```python
from datasets import load_dataset
import json
ds = load_dataset("anuj-inavlabs/kupe-tts", split="train")
row = ds[0]
print(row["text"])
print(row["audio"]["sampling_rate"], len(row["audio"]["array"]))
meta = json.loads(row["meta_json"])
chars = json.loads(row["characters_json"])
```

Loose `audio/` files may still exist for legacy access; **prefer these clusters**
for training and the Dataset Viewer (everything is inside the parquet).