wav2vec2-large-xls-r-300m-Urdu

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

• Loss: 0.9889
• Wer: 0.5607
• Cer: 0.2370

python eval.py --model_id kingabzpro/wav2vec2-large-xls-r-300m-Urdu --dataset mozilla-foundation/common_voice_8_0 --config ur --split test

Inference With LM

from datasets import load_dataset, Audio
from transformers import pipeline
model = "kingabzpro/wav2vec2-large-xls-r-300m-Urdu"
data = load_dataset("mozilla-foundation/common_voice_8_0",
                     "ur",
                     split="test", 
                     streaming=True, 
                     use_auth_token=True)

sample_iter = iter(data.cast_column("path", 
                    Audio(sampling_rate=16_000)))
sample = next(sample_iter)

asr = pipeline("automatic-speech-recognition", model=model)
prediction = asr(sample["path"]["array"], 
                  chunk_length_s=5, 
                  stride_length_s=1)
prediction
# => {'text': 'اب یہ ونگین لمحاتانکھار دلمیں میںفوث کریلیا اجائ'}

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 32
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• num_epochs: 200

Training results

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.2+cu102
• Datasets 1.18.2.dev0
• Tokenizers 0.11.0

Eval results on Common Voice 8 "test" (WER):