FilSwitch / README.md
qwerttyuiiop's picture
Update README.md
549058f verified
|
Raw
History Blame Contribute Delete
1.94 kB
metadata
task_categories:
  - automatic-speech-recognition
language:
  - tl
tags:
  - Audio

FilSwitch: A Filipino–English Code-Switched Speech Dataset

A Filipino-English (Taglish) read-speech dataset created for training and evaluating automatic speech recognition (ASR) systems on code-switched speech. The corpus contains 3,555 manually validated utterances from 152 speakers, totaling approximately 8.9 hours of audio.

Languages Filipino + English code switching
Domain Read, Philippine news
Utterances 3,555 (16 kHz sampling rate)
Train / Test 6.89 h / 2.04 h (speaker disjoint)

ASR Results

Model Pretrained WER FilSwitch Fine-tuning FLEURS Fine-tuning
Whisper Large v3 13.62% 9.24% 14.71%
MMS 26.41% 17.60% 27.36%
Qwen3-ASR 25.17% 20.13% 26.45%
Gemma 4 E4B 16.18%

Fine-tuning on FilSwitch produced relative WER reductions of 32.16% for Whisper, 33.36% for MMS, and 20.02% for Qwen3-ASR. In contrast, fine-tuning on Filipino FLEURS did not result in WER reductions on the FilSwitch test set.

The evaluation and text normalization script can be found at ./eval_script.py.

Intended Use

The dataset is intended for Filipino-English ASR fine-tuning, benchmarking, and error analysis.

The dataset and accompanying study are currently under publication. Please contact sqwerttyuiiop<at>gmail.com for collection methodology (paper). If you found this work useful, please consider citing it.