---
license: other
language:
- nan
- zh
task_categories:
- automatic-speech-recognition
- translation
tags:
- taiwanese-hokkien
- speech-to-text-translation
- synthetic-data
configs:
- config_name: audio
  data_files:
  - split: train
    path: audio/nan_based_on_zh/train/*.parquet
- config_name: corpus
  data_files:
  - split: train
    path: corpus/nan_based_on_zh.train.jsonl
---

# TATTS

Synthetic Taiwanese Hokkien speech paired with written Mandarin targets.  This
release is intended for **listening in Taiwanese Hokkien and writing Mandarin**
(S2TT); `text` is therefore the Mandarin training target, while `tailo_hant`
and `tailo_pinyin` describe the spoken content.  The labels were not replaced
during packaging.

This is the latest fully frozen and audited snapshot from the ongoing Hua/Tailo
TTS pipeline. New generated rows are published only after the same integrity
checks have completed, so a live partial batch is never mixed into this release.

## Release statistics

- Rows: 106,325
- Audio: 233.385 hours
- Unique Mandarin targets: 4,253
- Variants per target: 25: 4,253 targets
- Audio Parquet shards: 30
- Audio encoding: FLAC, PCM16, 16 kHz, mono

## Configs and columns

The `audio` config is directly loadable with Hugging Face Datasets and embeds
FLAC bytes.  Important columns are `audio`, `text`, `tailo_hant`,
`tailo_pinyin`, `lang`, `duration`, source/provenance identifiers, speaker
attributes, and augmentation parameters.

The `corpus` config contains the same public metadata without binary audio.
Its `audio_filepath` values are portable Parquet row locators such as
`audio/nan_based_on_zh/train/part-....parquet#row=N`; no cluster-local absolute
paths are published.

## Packaging validation

Every frozen source row had a unique ID and audio path, accepted pipeline
statuses, and a complete 25-variant target group.  Every FLAC was checked
against its recorded byte size and SHA256, fully decoded, confirmed non-silent
PCM16 at 16 kHz mono, and checked against metadata duration within 1.01 audio
samples.  The Parquet files copied from node-local staging were verified again
by SHA256 before publication readiness was marked.

This corpus contains synthetic and augmented speech.  Review the upstream data
and voice-reference licensing/provenance before redistribution or commercial
use; the package does not claim a permissive blanket license.