Wav2Vec2-Large-Robust ETRI Korean-English 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.

Data Information

Dataset Name: ETRI English Pronunciation of Korean Learners
Train Data: 14,305 samples
Valid Data: 1,590 samples
Test Data: 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: 0.2527
Validation Loss: 0.4532
Word Error Rate (WER) on Validation Set: 0.1617

Test Results

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

Word Error Rate (WER): 0.1223

Phoneme Data Example

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

Sample 1:

Provided Sentence: The one with the ribbon on its head
Correct Korean English Phonemes: dh ah w ah n w ih dh ax r ih b ah n ao n ih t s hh eh dd
Predicted Phonemes: d ah w ah n w ih dh ah r ih b ah n ao n ih ts hh eh dd

Training Logs

TensorBoard logs are available for detailed training analysis:

events.out.tfevents.1732529747.oem-WS-C621E-SAGE-Series.2265579.0
events.out.tfevents.1732573537.oem-WS-C621E-SAGE-Series.2265579.1

Use the following command to visualize logs:

tensorboard --logdir=./logs/