Wav2Vec-Bert-2.0-ngram-Urdu

This model is a fine-tuned version of facebook/w2v-bert-2.0 on the Urdu split of the Common Voice 17 dataset. The fine-tuned model is enhanced with the addition of an ngram language model that has also been trained on the same dataset. It achieves the following results on the evaluation set:

• Loss: 0.3681
• Wer: 0.2407

Usage Instructions

from transformers import AutoFeatureExtractor, Wav2Vec2BertModel
import torch
from datasets import load_dataset

dataset = load_dataset("hf-internal-testing/librispeech_asr_demo", "clean", split="validation")
dataset = dataset.sort("id")
sampling_rate = dataset.features["audio"].sampling_rate

processor = AutoProcessor.from_pretrained("UmarRamzan/w2v2-bert-ngram-urdu")
model = Wav2Vec2BertModel.from_pretrained("UmarRamzan/w2v2-bert-ngram-urdu")

# audio file is decoded on the fly
inputs = processor(dataset[0]["audio"]["array"], sampling_rate=sampling_rate, return_tensors="pt")
with torch.no_grad():
    outputs = model(**inputs)

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-06
• 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: 100
• num_epochs: 1
• mixed_precision_training: Native AMP

Framework versions

• Transformers 4.40.2
• Pytorch 2.2.1+cu121
• Datasets 2.19.1
• Tokenizers 0.19.1