wav2vec2-ksponspeech

This model is a fine-tuned version of Wav2vec2-large-xlsr-53 on the None dataset.
It achieves the following results on the evaluation set:

• WER(Word Error Rate) for Third party test data : 0.373

For improving WER:

• Numeric / Character Unification
• Decoding the word with the correct notation (from word based on pronounciation)
• Uniform use of special characters (. / ?)
• Converting non-existent words to existing words

Model description

Korean Wav2vec with Ksponspeech dataset.

This model was trained by two dataset :

• Train1 : https://huggingface.co/datasets/Taeham/wav2vec2-ksponspeech-train (1 ~ 20000th data in Ksponspeech)
• Train2 : https://huggingface.co/datasets/Taeham/wav2vec2-ksponspeech-train2 (20100 ~ 40100th data in Ksponspeech)
• Validation : https://huggingface.co/datasets/Taeham/wav2vec2-ksponspeech-test (20000 ~ 20100th data in Ksponspeech)
• Third party test : https://huggingface.co/datasets/Taeham/wav2vec2-ksponspeech-test (60000 ~ 20100th data in Ksponspeech)

Hardward Specification

• GPU : GEFORCE RTX 3080ti 12GB
• CPU : Intel i9-12900k
• RAM : 32GB

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0003
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 30
• mixed_precision_training: Native AMP

Framework versions

• Transformers 4.19.4
• Pytorch 1.11.0
• Datasets 2.2.2
• Tokenizers 0.12.1