wav2vec2-large-xls-r-300m-ur-cv8

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: 1.1443
Wer: 0.5677

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: 0.0001
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: 1000
num_epochs: 200

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.6269	15.98	400	3.3246	1.0
3.0546	31.98	800	2.8148	0.9963
1.4589	47.98	1200	1.0237	0.6584
1.0911	63.98	1600	0.9524	0.5966
0.8879	79.98	2000	0.9827	0.5822
0.7467	95.98	2400	0.9923	0.5840
0.6427	111.98	2800	0.9988	0.5714
0.5685	127.98	3200	1.0872	0.5807
0.5068	143.98	3600	1.1194	0.5822
0.463	159.98	4000	1.1138	0.5692
0.4212	175.98	4400	1.1232	0.5714
0.4056	191.98	4800	1.1443	0.5677

Framework versions

Transformers 4.16.0
Pytorch 1.10.0+cu111
Datasets 1.18.1
Tokenizers 0.11.0

Evaluation Commands

To evaluate on mozilla-foundation/common_voice_8_0 with split test

python eval.py --model_id anuragshas/wav2vec2-large-xls-r-300m-ur-cv8 --dataset mozilla-foundation/common_voice_8_0 --config ur --split test

Inference With LM

import torch
from datasets import load_dataset
from transformers import AutoModelForCTC, AutoProcessor
import torchaudio.functional as F
model_id = "anuragshas/wav2vec2-large-xls-r-300m-ur-cv8"
sample_iter = iter(load_dataset("mozilla-foundation/common_voice_8_0", "ur", split="test", streaming=True, use_auth_token=True))
sample = next(sample_iter)
resampled_audio = F.resample(torch.tensor(sample["audio"]["array"]), 48_000, 16_000).numpy()
model = AutoModelForCTC.from_pretrained(model_id)
processor = AutoProcessor.from_pretrained(model_id)
input_values = processor(resampled_audio, return_tensors="pt").input_values
with torch.no_grad():
    logits = model(input_values).logits
transcription = processor.batch_decode(logits.numpy()).text
# => "اب نے ٹ پیس ان لیتے ہیں"

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

Without LM	With LM (run `./eval.py`)
52.146	42.376