barockok/dataset-alpha-v1

Dataset Alpha V1

A multilingual audio-text parallel corpus compiled from open-source phonetic research datasets.

Dataset Structure

The dataset is distributed as a split archive. Reassemble with:

cat dataset-alpha-v1.tar.gz.part* > dataset-alpha-v1.tar.gz
tar xzf dataset-alpha-v1.tar.gz

Subsets

Subset	Samples	Hours	Source Type	Language Family
A	2,921	~7.5	Parallel speech-text (read speech)	Austronesian, ~200M speakers
B	5,822	~9.5	Regional speech-text (read speech)	Austronesian, ~80M speakers
C	4,213	~4.5	Regional speech-text (read speech)	Austronesian, ~40M speakers
Total	12,956	21.5

Format

Each subset contains:

• *.wav — stereo audio at 24kHz (left channel = primary, right channel = secondary)
• *.json — word-level timestamp alignments: {"alignments": [[word, (start, end), speaker], ...]}
• manifest.jsonl — per-file metadata: {"path": "...", "duration": seconds}

All transcriptions use Latin script with language-appropriate diacritics.

Iterative Updates

This dataset is actively growing. Subsets and metadata are updated as new sources are added.

Usage

from datasets import load_dataset

ds = load_dataset("barockok/dataset-alpha-v1")