Hugging Face's logo

Datasets:
pelatihpokemongo
/
datatalk

Tasks:
Automatic Speech Recognition
Text-to-Speech
Modalities:
Audio
Text
Formats:
parquet
Languages:
Indonesian
Size:
1K - 10K
Tags:
audio
speech
tts
asr
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
datatalk / README.md
pelatihpokemongo's picture
pelatihpokemongo
Fix dataset card header (remove configs/data_files auto)
9cb9ff3 verified
preview code
|
raw
history blame contribute delete
2.89 kB
metadata 
pretty_name: datatalk (Indonesian)
tags:
  - audio
  - speech
  - tts
  - asr
language:
  - id
license: cc-by-4.0
task_categories:
  - automatic-speech-recognition
  - text-to-speech
size_categories:
  - 1K<n<10K

pelatihpokemongo/datatalk

An Elise-style Indonesian speech dataset with two columns: audio and text, published for quick prototyping in ASR/TTS. Built from local audio and transcripts with a single train split. Viewer is auto-converted to Parquet for fast preview on the Hub.

Dataset summary

  • Split: train
  • Rows: 5,400
  • Columns:
    • audio (datasets.Audio): waveform automatically decoded by the Hub
    • text (string): transcript in Indonesian

Intended use

  • Quick experiments for ASR baselines or TTS tokenizers/aligners.
  • Finetuning demos where a compact Indonesian set with transcripts is useful.

How it was built

Data were collected from /Users/benny/Development/side/dataset/audio/id and assembled using the helper script scripts/build_hf_dataset.py in this workspace. Key rules:

  • Detect transcripts from a manifest if present (e.g., validated.tsv, CSV/TSV/JSONL/JSON); otherwise look for sidecar .txt files (same basename).
  • Random sampling with seed 42 to pick the target number of rows.
  • Enforced non-empty transcripts (--no-empty-text).
  • Cast audio column to datasets.Audio() so it is playable in the Viewer and durations are computed.
  • Uploaded using Dataset.push_to_hub("pelatihpokemongo/datatalk").

Repro command used most recently:

python scripts/build_hf_dataset.py \
  --data-root /Users/benny/Development/side/dataset/audio/id \
  --repo-id pelatihpokemongo/datatalk \
  --num-rows 5400 \
  --seed 42 \
  --no-empty-text

Usage 

from datasets import load_dataset

ds = load_dataset("pelatihpokemongo/datatalk", split="train")
example = ds[0]
audio = example["audio"]  # dict with 'array', 'sampling_rate', 'path'
text = example["text"]
print(text)

Streaming preview (large repos) is also supported:

ds_stream = load_dataset("pelatihpokemongo/datatalk", split="train", streaming=True)
first = next(iter(ds_stream))

License

This dataset card declares cc-by-4.0. Ensure your upstream audio/transcript sources are compatible with the selected license. If you need a different license, update this card accordingly.

Limitations

  • Compact size; not representative of all Indonesian accents or domains.
  • Text normalization and punctuation are minimally curated.

Acknowledgements

  • Thanks to the Hugging Face Datasets team for the excellent tooling.
  • Structure inspired by MrDragonFox/Elise.