Wav2Vec2-Large-Robust finetuned on the revised ETRI Data of Korean-English for Pronunciation Model

This repository contains a fine-tuned Wav2Vec2-Large-Robust model for phoneme recognition tasks. The model was trained and evaluated on our in-house English pronunciations of Korean learners dataset, which was made with ETRI and revised by SNU SLP lab.

Data Information

Dataset Name: English Pronunciation of Korean Learners (made with ETRI) revised by SNU SLP lab
Data Type: Speech recordings of Korean learners speaking English, annotated with phoneme sequences.
Annotation: Each utterance is transcribed at the phoneme level, including pronunciation errors marked with _err. These errors highlight phoneme substitutions, insertions, and deletions that occur due to the influence of the Korean language on English pronunciation.
Train Set: 14,305 samples
Valid Set: 1,590 samples
Test Set: 3,974 samples

Training Procedure

The model was fine-tuned for phoneme recognition using the Hugging Face transformers library. Below are the training steps:

Data preprocessing to align audio with phoneme labels.
Wav2Vec2-Large-Robust model fine-tuning with CTC loss.
Evaluation on validation and test sets.

Training Hyperparameters

Epochs: 50
Learning Rate: 0.0001
Warmup Ratio: 0.1
Scheduler: Linear
Batch Size: 8
Loss Reduction: Mean
Feature Extractor Freeze: Enabled

Training Results

The following metrics were achieved during training:

Final Training Loss: 0.2415
Phoneme Error Rate (PER) on Training Set: 0.0508
Validation Loss: 0.3999
Phoneme Error Rate (PER) on Validation Set: 0.1622

Test Results

The model was evaluated on the test dataset with the following performance:

Phoneme Error Rate (PER): 0.0905

Phoneme Data Example

Below is an example of how the dataset is structured for phoneme recognition tasks:

Sample :

Provided Sentence: I'M LOOKING FOR MY PUPPY LUKE HE RAN AWAY THIS MORNING
True Phonemes of Korean pronunciation: ay m l uh k ih_err ng f er m ay p ah p iy l uw k hh iy l ah n ah w ey dh ih s m ao r n ih_err n
Predicted Phonemes: ay m l uh k ih ng f er m ay p ah p iy l uh k hh iy l ah n ah w ey dh ih s m ao r n ih_err ng

Training Logs

TensorBoard logs are available for detailed training analysis:

events.out.tfevents.1737043507.oem-WS-C621E-SAGE-Series.1534005.0
events.out.tfevents.1737088179.oem-WS-C621E-SAGE-Series.1534005.1

Use the following command to visualize logs:

tensorboard --logdir=./logs/