Philippine speech AI · research log

Training Filipino speech models for local use.

We train on GB10, measure what changed, and work toward transcription that runs on a Mac. Every result below has a defined dataset and scoring protocol.

Read the training study Download the results

Completed experiments

Results, including the gaps.

A17 · corrected Whisper Small training

The pipeline learned. Longer training did not close the gap.

A three-rate LoRA sweep followed by a fresh 4,000-step run. Final WER was 31.02%, compared with 30.95% for the short reference. A transient spike recovered, but its cause remains unknown.

4,510 training clips9 held-out speakersOne seed

Read Study 01

A25 · existing models, shared audio

Large delivered better recognition at a higher processing cost.

Four existing models, 200 development recordings, and three audio conditions. Large v3 PLD normalized had the lowest WER and CER in every condition, at about four times Small’s median processing time on GB10.

2,400 successful transcriptionsClean, noise, dropoutInference comparison

Read Study 02

Next release · in preparation

A model for the Mac.

Our next candidate will use reviewed data, then face matched quality checks and end-to-end Mac evaluation. The model and curated dataset are planned for JC’s Hugging Face account.

See the release requirements

Working name · Suman Ube 1

October 18 is the requested release target. The next training run, reviewed dataset, and Mac artifact are pending. No quality improvement or model download is claimed yet.