Wav2VecBert 2.0 Khmer

This model is a fine-tuned version of facebook/w2v-bert-2.0 on the OpenSLR 42 dataset.

from transformers import pipeline
recognizer = pipeline("automatic-speech-recognition", model="seanghay/w2v-bert-2.0-khmer", device="cuda")
text = recognizer("audio.mp3", chunk_length_s=10, stride_length_s=(4, 2))["text"]

Training and evaluation data

25.79% WER (Eval with 10% of OpenSLR seed: 42)

{
  "epoch": 14.634146341463415,
  "eval_loss": 0.36365753412246704,
  "eval_runtime": 8.7546,
  "eval_samples_per_second": 33.24,
  "eval_steps_per_second": 4.226,
  "eval_wer": 0.2579008973858759,
  "step": 2400
}

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 15

Framework versions

• Transformers 4.42.4
• Pytorch 2.0.1
• Datasets 2.20.0
• Tokenizers 0.19.1