| --- |
| language: |
| - hak |
| task_categories: |
| - automatic-speech-recognition |
| - text-to-speech |
| pretty_name: Hakka TTS Synthetic Speech Corpus |
| size_categories: |
| - 1M<n<10M |
| --- |
| |
| # Hakka TTS Synthetic Speech Corpus |
|
|
| This public corpus contains 6,498,285 synthetic Hakka utterances |
| (15,539.41 hours) as 16 kHz FLAC audio. The FLAC files are stored in |
| uncompressed tar shards; FLAC itself provides lossless audio compression. |
|
|
| ## Layout |
|
|
| - `metadata.jsonl`: one row per audio file. |
| - `audio_shards/audio-NNNNN.tar`: FLAC archive shards. |
| - `package_stats.json`: package counts and sizes. |
|
|
| Each metadata row contains `text`, `hakka_text`, `lang`, `duration`, |
| `audio_archive`, and `audio_filepath`. To resolve one row, open the relative |
| tar file named by `audio_archive` and read the member named by |
| `audio_filepath`. If all tar files are extracted at the repository root, |
| `audio_filepath` is also the extracted relative path. |
|
|
| The `source` field distinguishes `moe`, `elearning`, and generated `concat` |
| records. The `dialect` field preserves the dialect directory assigned by the |
| generation pipeline. |
|
|
| ## Integrity |
|
|
| The package is published only after every input row is written exactly once. |
| Archive members and metadata rows are generated in the same loop, and the |
| final row count must match 6,498,285 before upload starts. |
|
|
