sapinsapin/filipinospeechcorpus

Filipino Speech Corpus

Word and sentence-level segments from the Filipino Speech Corpus (FSC), stored as a Hugging Face Parquet dataset with raw 16kHz mono audio. For more detailed information about the data, checkout the full paper Development of a Filipino Speech Corpus.

The corpus is composed of single-word utterances, read speech, and spontaneous speech from 100 speakers aged 16 and above.

Citation

If you use this dataset, please cite the original corpus:

@article{sagumdevelopment,
  title={DEVELOPMENT OF A FILIPINO SPEECH CORPUS},
  author={Sagum, Ramil}
}

Usage

Whisper / ASR:

from datasets import load_dataset, Audio
ds = load_dataset("sapinsapin/filipinospeechcorpus")
ds = ds.filter(lambda x: x["num_words"] >= 3 and x["duration"] >= 1.5)
ds = ds.cast_column("audio", Audio(sampling_rate=16000))

TTS (LJSpeech-compatible, 22050Hz):

ds = ds.filter(lambda x: x["speech_type"] == "read" and 1.0 <= x["duration"] <= 10.0)
ds = ds.cast_column("audio", Audio(sampling_rate=22050))

Schema

Field	Type	Description
audio	Audio(16000)	16kHz mono WAV segment
sentence	str	Transcription
duration	float	Segment duration in seconds
num_words	int	Word count
speaker_id	str	Speaker identifier
gender	str	male / female
age_group	str	Age range e.g. 20-27
speech_type	str	read / spontaneous / machine
source_file	str	Original TRS stem

File naming convention

The names of the sound files have the following information: speaker identification number, speaker gender, speaker age group and text material set. For example, the file name, 12-0-4-2.wav, refers to 12 --speaker id no.

0--gender 0-male 1-female

4-age group 0:20-27 1:28-35 2:36-43 3:44-51 4:52-60

2-text material number

Code

Processing code is available at https://github.com/sapinsapin/halohalo

Splits

Split	Rows
train	90%
test	10%