LauelKills/sugidanon-hil-codeswitch

Sugidanon 🎙️ — Code-Switch Hiligaynon Speech Benchmark

The first openly-licensed, code-switch-labeled speech benchmark for Hiligaynon (Ilonggo) — a language spoken by 9M+ Filipinos yet nearly invisible to modern speech technology.

 ·  Code: https://github.com/Jazztinn/tinig-sa-liwanag  ·  License: CC BY 4.0

Real Hiligaynon-English-Tagalog speech, labeled per word, with a scorer that measures what generic models miss: accuracy at the moment the language switches. Run the one-click Colab to reproduce the headline result on a fresh machine in minutes.

Why it matters

The Philippines has 130+ languages. Tagalog ASR has advanced; regional tongues like Hiligaynon have no open speech datasets, benchmarks, or models. Real Ilonggo speech constantly mixes Hiligaynon, Tagalog, and English — exactly where off-the-shelf systems break. Sugidanon makes that failure measurable, so the next developer has a building block instead of a blank page.

Headline result

Whisper small (--language tl) over the 40 clips:

Metric	WER
Overall	59.8%
Monolingual (Hiligaynon)	66.3%
Switch-region	36.4%
Switch penalty	−30.0%

By pair: hil↔en 40.8%, hil↔tl 24.4%, tl↔en 6.2%.

The negative penalty is the finding: a Tagalog model nails the borrowed English/Tagalog words but fails on the Hiligaynon matrix it was never trained on. tl↔en is near-solved (6%); hil↔en is worst (41%). The gap scales with Hiligaynon — precisely what this dataset exists to expose.

What's inside

80+ code-switch utterances recorded by Aziel Faith Agustin, a Hiligaynon (Ilonggo) speaker who also reviewed the sentences, across 8 everyday domains (market, transport, school/work, family, health, culture, everyday, oral tradition / heritage) and 4 switch types (HIL, HIL+EN, HIL+TL, HIL+TL+EN). Every word carries a hil/tl/en tag, so the scorer can isolate switch-region errors.

Transcripts were reviewed by the speaker; per-word language tags are auto-seeded (lang_tags_status: seed_unverified) pending a confirmation pass. Single speaker — a seed benchmark to extend, not a final model-ranking corpus.

Quick start

from datasets import load_dataset
ds = load_dataset("LauelKills/sugidanon-hil-codeswitch", data_dir="data/audio", split="train")
ds[0]["audio"], ds[0]["transcript"], ds[0]["switch_type"], ds[0]["tokens"]

Or reproduce the benchmark end-to-end with the one-click Colab badge above.

Languages

• hil: Hiligaynon / Ilonggo
• tl: Filipino / Tagalog
• en: English
• other: proper nouns, unclear tokens, or other languages

Dataset Structure

Expected repository layout:

data/
  audio/
    <clip_id>.wav
  annotations/
    <clip_id>.json
  predictions/
    asr/
      whisper-large-v3-tl/
        <clip_id>.json
      mms-1b-all/
        <clip_id>.json

Each annotation contains:

• clip ID
• audio path
• duration
• coarse speaker metadata
• matrix language
• tokenized transcript
• per-token language labels

Example:

{
  "clip_id": "hil_cs_001",
  "audio_file": "audio/hil_cs_001.wav",
  "duration_sec": 3.39,
  "domain": "market",
  "switch_type": "HIL+EN",
  "transcript": "Pila ang grocery budget naton para sa weekend?",
  "matrix_language": "hil",
  "review_status": "reviewed",
  "lang_tags_status": "seed_unverified",
  "tokens": [
    { "idx": 0, "text": "Pila", "lang": "hil" },
    { "idx": 1, "text": "ang", "lang": "hil" },
    { "idx": 2, "text": "grocery", "lang": "en" },
    { "idx": 3, "text": "budget", "lang": "en" },
    { "idx": 4, "text": "naton", "lang": "hil" }
  ]
}

Evaluation

One-command reproduction:

python3 scripts/eval_asr_baselines.py

The script evaluates every model directory under:

data/predictions/asr/

It reports:

• overall WER
• switch-region WER
• monolingual WER
• switch penalty

Lower-level scoring:

python3 score.py --ref data/annotations --hyp data/predictions/asr/whisper-large-v3-tl

Baselines

Included worked-example prediction directories:

• whisper-large-v3-tl
• mms-1b-all

These files demonstrate the evaluation format. They are not final published benchmark numbers until real model outputs over the full test set are added.

Annotation Guidelines

See:

docs/transcription_guidelines.md

Data protection & ethics

These are human voice recordings — personal data — handled accordingly.

• Consent. Recorded only from speakers who agreed to open release (CC BY 4.0) and reuse for speech-technology research, obtained before recording.
• Minimal / anonymized PII. Scripted elicited sentences, not private conversations. Speakers are identified by a coarse anonymized id (e.g. spk01) plus optional non-identifying metadata (region, age band, gender, Hiligaynon fluency). No contact details or exact location in the data; a name appears only as voluntary credit.
• Right to withdraw. A speaker may request removal; affected audio and annotations are deleted from the dataset.
• Honest labeling. speaker.fluency is recorded (native / fluent / non_native); non-native clips are flagged and must not be used as native gold data. AI-assisted tags are seed_unverified until human-reviewed.
• Intended use. A research/evaluation building block for inclusive Philippine speech technology. Do not use it to identify, profile, surveil, or impersonate speakers, or to build voice-cloning systems without the speakers' separate explicit consent. If you reuse it: preserve attribution, honor withdrawal requests, and do not attempt to re-identify speakers.

Licensing

Dataset files created by Team Hague are released under CC BY 4.0.

Code is released under MIT.

Third-party resources are not relicensed. See:

docs/licensing.md
LICENSE

Acknowledgments

This dataset exists thanks to Aziel Faith Agustin, the Hiligaynon (Ilonggo) speaker who reviewed the elicitation sentences and recorded all 40 clips. The reference transcripts and audio are their voice and review.

Citation

@dataset{team_hague_sugidanon_2026,
  title = {Sugidanon: Code-switched Hiligaynon speech and translation benchmark scaffold},
  author = {{Team Hague}},
  note = {Speech recorded and reviewed by Aziel Faith Agustin (Hiligaynon speaker)},
  year = {2026},
  license = {CC BY 4.0}
}

Baseline result

Whisper small (--language tl) over the 40 clips:

Metric	WER
Overall	61.4%
Monolingual (Hiligaynon)	69.2%
Switch-region	38.8%
Switch penalty	−30.4%

The negative penalty is the finding: an off-the-shelf Tagalog model handles the borrowed English/Tagalog switch words but fails on the Hiligaynon matrix. See docs/evaluation_report.md. Preliminary (Whisper small, single speaker).