HandsFree Whisper-tiny (Filipino/Tagalog)

Fine-tuned openai/whisper-tiny for Filipino and Taglish voice commands. Undergraduate thesis, Holy Angel University, School of Computing, 2026.

Results

Model WER
Baseline openai/whisper-tiny 82.72%
This model (best checkpoint, step 750) 63.15%

23.7% relative reduction, evaluated on the same 1,479-example held-out test set with identical evaluation settings.

Training data

6,227 train / 1,479 test examples combining:

  • Filipino Speech Corpus (sapinsapin/filipinospeechcorpus), filtered to human speech only, min 3 words, min 1.5s duration
  • Google FLEURS fil_ph

Training

Full fine-tune, HuggingFace Seq2SeqTrainer, 2000 steps, lr 1e-5, effective batch 16, fp16, single RTX 3050.

Downloads last month
-
Safetensors
Model size
37.8M params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support