sapinsapin/filipinospeechcorpus

Filipino Speech Corpus (FSC)

Studio-recorded Filipino read, spontaneous, and word-level speech — 125 speakers, packaged as ready-to-stream Parquet.

<div align="center">

313,322 transcribed segments · 65.1 hours · 125 speakers · 16kHz mono

  

</div>

This is the Filipino Speech Corpus (Sagum), recorded in a controlled setting and hand/machine transcribed with Transcriber. This repo repackages the original .wav + .trs volumes as segment-level Parquet with inline audio, so you can stream it without downloading and parsing XML.

New here? Start with the [30-second quickstart](#30-second-quickstart), then read [Before you train](#-before-you-train) — the segment length distribution will surprise you.

30-second quickstart

from datasets import load_dataset

ds = load_dataset("sapinsapin/filipinospeechcorpus", split="train", streaming=True)
row = next(iter(ds))

print(row["sentence"], row["duration"], row["speech_type"])
# audio arrives decoded as a numpy array at 16kHz

Full download (6.9 GB):

ds = load_dataset("sapinsapin/filipinospeechcorpus")   # train + test

⚠️ Before you train

This corpus is mostly single words, not sentences. The median segment is 0.62 seconds and 95% of segments are under 1.3 s, because 56% of the corpus is machine-pre-segmented word tokens intended for unit-selection synthesis and keyword work. If you load it and train directly, you will train on isolated words.

For sentence-level ASR or TTS you want the read/spontaneous portion with a length filter — which leaves roughly 8,500 utterances (~7 hours):

ds = ds.filter(lambda x:
    x["speech_type"] in ("read", "spontaneous")
    and 1.5 <= x["duration"] <= 30.0
    and x["num_words"] >= 3
)

Three more things worth knowing before you spend GPU hours:

What's inside

All statistics on this card are computed from the source corpus transcriptions (313,322 segments); 305,246 of those survive into the published Parquet — the rest are empty turns, control markers, or turns whose audio is missing.

Splits (published rows)

Speech types — the three source volumes, and the reason the length distribution is bimodal:

Duration

Speakers — 125, near-balanced by gender (51.8% male / 48.2% female of segments). Age skews hard young: 97.5% in 20-27, 1.6% in 28-35, 0.9% in 36-43.

Schema

Speaker metadata is decoded from the corpus filename convention (09_xx00xxxx_15A → speaker 09, male, age band 20–27, session 1, set 5A).

Models trained on this data

Reference finetunes, trained by the same pipeline, so you can hear/measure what the corpus supports before committing to your own run:

Both are demonstration baselines on a single 8 GB GPU, not state-of-the-art — they exist to prove the data path end to end and to give you a known-good starting configuration.

Reproduce either in one command:

python finetune_tts.py --dataset fsc --push
python finetune_asr.py --dataset fsc --push

How it was built

1. 1.Parse Transcriber .trs XML from Volume 6 (Transcriptions) — three directories, one per speech_type.
2. 2.Slice each turn out of the corresponding .wav at its annotated start/end, resample to 16 kHz mono.
3. 3.Drop empty turns and Transcriber control markers (.., {...} events).
4. 4.Decode speaker/gender/age from the filename convention.
5. 5.Shard to Parquet with audio inline, 90/10 random split.

Of 313,322 transcribed segments in the source, 305,246 survive; the remainder are empty turns, control markers, or turns whose audio is missing.

Pipeline source: `process_fsc_parquet.py`

Limitations

• —Not a sentence corpus. See Before you train.
• —Read speech is scripted. Prosody reflects reading, not conversation.
• —Age and register are narrow — young adults, mostly in one setting.
• —Transcription conventions vary between the hand-transcribed and machine-pre-segmented volumes; the machine volume is word-aligned output, not editorial transcription.
• —No dialect labels. The corpus is Filipino/Tagalog; regional variation is not annotated.
• —Original recording notes flag per-speaker irregularities (e.g. a wrong prompt set given to speaker 66, and several speakers withdrawn from the corpus).

Related datasets

Part of the halohalo Philippine-language speech family:

License, source and citation

The recordings originate from the Filipino Speech Corpus developed by Ramil Sagum. This repackaging is distributed for research use — cite the original corpus, not this repo:

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

Paper: Development of a Filipino Speech Corpus

If you represent the corpus authors and want the terms or attribution changed, please open a discussion on this repo.

Contributing

Philippine languages are under-served in speech ML, and this family is built in the open so others can pick it up. Useful contributions:

• —Report bad segments via the Community tab (include source_file)
• —Share finetunes trained on it — tag sapinsapin/filipinospeechcorpus in your model card and it will appear in this repo's "used by" list
• —Improve the pipeline: github.com/sapinsapin/halohalo