README.md

---
configs:
- config_name: ja
  data_files:
  - split: train
    path: corpus/ja.train.clean.jsonl
---

# ja high-quality speech corpus

This dataset contains the guarded high-quality Japanese speech corpus and its source-partitioned audio.

All audio is FLAC PCM16. Source sample rate and channel count are preserved. Every packaged item passed a source-to-FLAC bit-identical PCM round trip and a separate full-release decode/hash validation.

Corpus rows use portable locators of the form `audio/<subset>/<group>/part-N.parquet#row=N`.

## Corpus overview

| Corpus/config | Rows | Approx. hours |
|---|---:|---:|
| ja | 20,354,333 | 35,729.01 |

## Audio subset overview

Counts below are unique packaged audio rows.

| Audio subset | Unique rows | Approx. hours | Rejected rows | Parquet GiB |
|---|---:|---:|---:|---:|
| game_dialogue | 3,922,808 | 6,211.25 | 1 | 865.84 |
| anime_dialogue | 65,438 | 101.92 | 0 | 14.47 |
| anime_dialogue_v2 | 230,170 | 377.90 | 0 | 48.99 |
| character_dialogue | 4,469 | 6.79 | 0 | 0.73 |
| general_speech | 16,131,448 | 29,031.39 | 0 | 1,872.31 |

## JSONL examples

Two abridged metadata examples are shown below. Physical audio locators and source identifiers are visible only after access approval.

```jsonl
{"text":"うぅ～ん。","context":"","duration":1.58,"lang":"<|JA|>","language":"<|JA|>"}
{"text":"世界、大丈夫？","context":"うぅ～ん。","duration":2.86,"lang":"<|JA|>","language":"<|JA|>"}
```

## Download

Approved users can download the complete gated repository:

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="WTForbes/ja",
    repo_type="dataset",
)
```

Or load the corpus configuration:

```python
from datasets import load_dataset
corpus = load_dataset("WTForbes/ja", "ja")
```

## Release totals

- Unique audio: 20,354,333
- Audio hours: 35,729.247
- Parquet size: 3.008974 TB (2.736646 TiB)

See `release.json` for exact per-subset counts, hours, bytes, and corpus checksums.