whisper-small-taglish-fine-tuned-ct2

whisper-small-taglish-fine-tuned-ct2

A fine-tune of openai/whisper-small for Taglish (Tagalog-English code-switched) speech, converted to CTranslate2 format for fast inference with faster-whisper.

It was developed as part of Rekapo, a bilingual Filipino-English near real-time meeting transcription and summarization system (undergraduate thesis, ACLC College of Taytay).

Why Taglish?

Tagalog is already a low-resource language for speech recognition, and Taglish (Tagalog-English code-switching) is scarcer still. Whisper has a Tagalog language token but no mode for code-switching, so mixed-language speech is forced into a single language setting, and the minority language in a sentence tends to get mangled. Labeled Taglish data and public benchmarks are also hard to find. This model targets that gap.

Results

Word error rate (WER) on Taglish speech, lower is better:

That is a 14-point absolute reduction, or a 29% relative reduction in WER.

Baseline decoding setting: [language="tl" / "en" / auto-detect. State which one was used for both models.]

Evaluation set: [describe the test set: source, number of clips, total duration, speaker/domain mix] Text normalization: [e.g. lowercased, punctuation stripped]

Usage

pip install faster-whisper

from faster_whisper import WhisperModel

model = WhisperModel(
    "michaelildefonso/whisper-small-taglish-fine-tuned-ct2",
    device="cpu",          # or "cuda"
    compute_type="int8",   # "float16" on GPU
)

segments, info = model.transcribe("audio.wav", language="tl", beam_size=5)

for segment in segments:
    print(f"[{segment.start:.2f}s -> {segment.end:.2f}s] {segment.text}")

Notes:

• language="tl" works well for mixed Tagalog-English speech. Setting language=None enables auto-detection, which can misfire on heavily code-switched audio.
• Add word_timestamps=True for word-level timings.
• Audio is resampled to 16 kHz mono internally.

Training

• Base model: openai/whisper-small
• Training data: [describe the dataset: sources, hours, licensing, how it was curated]
• Checkpoint: step 3200
• Training setup: [epochs, learning rate, batch size, hardware]
• Conversion: fine-tuned checkpoint converted to CTranslate2 for faster-whisper

Intended use

Transcribing conversational Taglish speech such as meetings, interviews, and casual conversation.

Limitations

• Accuracy is best on speech similar to the training data. Noisy audio, strong accents, overlapping speakers, and specialized vocabulary will increase errors.
• Spelling of Tagalog and code-switched words is not standardized, so some "errors" are valid alternative spellings. WER can overstate the true error rate.
• Not evaluated for languages other than Tagalog and English.
• Not suitable for high-stakes transcription (legal, medical) without human review.

Demo

[Link to your Space or demo page once it's up]

Citation

If you use this model, please cite it as:

@misc{ildefonso2026whispertaglish,
  author = {Ildefonso, Michael},
  title  = {whisper-small-taglish-fine-tuned-ct2},
  year   = {2026},
  publisher = {Hugging Face},
  url    = {https://huggingface.co/michaelildefonso/whisper-small-taglish-fine-tuned-ct2}
}

Acknowledgements

Built on OpenAI's Whisper and the CTranslate2 / faster-whisper projects.