sapinsapin/filipinospeechcorpus

Filipino Speech Corpus (FSC)

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

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

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, then read 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:

Gotcha	Detail	What to do
Extreme outliers	Longest segment is 1,640 s (27 min); shortest is ~0 s	Always bound duration
Speaker overlap	Split is a random 90/10 over segments, so speakers appear in both	Re-split on speaker_id for speaker-disjoint eval
Narrow demographics	97.5% of segments come from the 20–27 age band	Don't claim age robustness

---

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)

Split	Segments
train	274,730
test	30,516
total	305,246

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

speech_type	Segments	Share	Speakers	What it is
machine	175,854	56.1%	64	Machine pre-segmented word tokens from read speech (Volume 6)
read	130,001	41.5%	50	Hand-transcribed read speech — paragraphs, sentences, word lists
spontaneous	7,467	2.4%	65	Hand-transcribed free speech (Volume 5)

Duration

Median	0.62 s
Mean	0.75 s
p95	1.21 s
Min / Max	~0 s / 1,639.6 s
Total	65.1 h

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

Field	Type	Description
audio	Audio(16000)	16 kHz mono segment, decoded on access
sentence	str	Transcription as written by the annotator
duration	float	Segment length in seconds
num_words	int	Whitespace word count
speaker_id	str	Speaker number from the filename
gender	str	male / female / unknown — recorded by observation
age_group	str	20-27, 28-35, 36-43, 44-51, 52-60, unknown
speech_type	str	read / spontaneous / machine
source_file	str	Original .trs stem, for tracing back to the corpus

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:

Model	Task	Base	Notes
sapinsapin/speecht5_tts-fsc	Text-to-speech	microsoft/speecht5_tts	1,000 steps on 1,867 read clips · eval loss 0.443 · listen to samples
sapinsapin/whisper-small-fsc	Speech recognition	openai/whisper-small	2,000 steps on 10k clips · WER 15.9% · CER 7.1% on the held-out split

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

Dataset	What it covers	Scale
filipinospeechcorpus (this one)	Filipino studio read + spontaneous speech	305k segments · 65 h
sapinsapin/pld	10 Philippine languages, prompted recordings	334k utterances · 448 h
sapinsapin/halo-livestream	Taglish code-switched livestream speech	seed release

---

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