wav2vec2-large-xls-r-300m-tr-cv16.1

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the common_voice_16_1 dataset. It achieves the following results on the evaluation set:

• Loss: 0.3356
• Wer: 0.4160

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0003
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 2
• mixed_precision_training: Native AMP

Model Inference

from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

model = Wav2Vec2ForCTC.from_pretrained("rumeyskeskn/wav2vec2-large-xls-r-300m-tr-cv16.1").to("cpu")
processor = Wav2Vec2Processor.from_pretrained("rumeyskeskn/wav2vec2-large-xls-r-300m-tr-cv16.1")
audio_path = "audio.wav"

audio_array, sampling_rate = librosa.load(audio_path, sr=16000)

input_values = processor(audio_array, sampling_rate=sampling_rate).input_values[0]

input_dict = processor(input_values, return_tensors="pt", padding=True)


logits = model(input_dict.input_values).logits

pred_ids = torch.argmax(logits, dim=-1)
prediction = processor.decode(pred_ids[0])

print("Prediction:")
print(prediction)

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
5.669	0.39	400	1.2228	0.8840
0.6809	0.78	800	0.6371	0.6557
0.4224	1.17	1200	0.4607	0.5226
0.3151	1.56	1600	0.3671	0.4457
0.2633	1.95	2000	0.3356	0.4160

Framework versions

• Transformers 4.38.2
• Pytorch 2.2.1+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2