metadata

license: apache-2.0
tags:
  - generated_from_trainer
model-index:
  - name: wav2vec2-xls-r-phone-mfa_korean
    results: []
language:
  - ko
metrics:
  - wer
pipeline_tag: automatic-speech-recognition

wav2vec2-xls-r-300m_phoneme-mfa_korean

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on a phonetically balanced native Korean read-speech corpus.

Training and Evaluation Data

Training Data

Data Name: Phonetically Balanced Native Korean Read-speech Corpus
Num. of Samples: 54,000 (540 speakers)
Audio Length: 108 Hours

Evaluation Data

Data Name: Phonetically Balanced Native Korean Read-speech Corpus
Num. of Samples: 6,000 (60 speakers)
Audio Length: 12 Hours

Training Hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0001
train_batch_size: 8
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 2
total_train_batch_size: 16
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_ratio: 0.2
num_epochs: 20 (EarlyStopping: patience: 5 epochs max)
mixed_precision_training: Native AMP

Evaluation Results

Phone Error Rate 3.88%
Monophthong-wise Error Rates: (To be posted)

Output Examples

MFA-IPA Phoneset Tables

Vowels

Consonants

Experimental Results

Official implementation of the paper (ICPhS 2023)
Major error patterns of L2 Korean speech from five different L1s: Chinese (ZH), Vietnamese (VI), Japanese (JP), Thai (TH), English (EN)

Framework versions

Transformers 4.21.3
Pytorch 1.12.1
Datasets 2.4.0
Tokenizers 0.12.1