anuragshas/wav2vec2-large-xls-r-300m-pa-in

XLS-R-300M - Punjabi

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: 1.2548
• 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: 0.0003
• 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_ratio: 0.12
• num_epochs: 120
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
6.4804	16.65	400	1.8461	1.0
0.474	33.33	800	1.1018	0.6624
0.1389	49.98	1200	1.1918	0.6103
0.0919	66.65	1600	1.1889	0.6058
0.0657	83.33	2000	1.2266	0.5931
0.0479	99.98	2400	1.2512	0.5902
0.0355	116.65	2800	1.2548	0.5677

Framework versions

• Transformers 4.15.0
• Pytorch 1.10.0+cu111
• Datasets 1.18.0
• Tokenizers 0.10.3

Evaluation Commands

1. To evaluate on mozilla-foundation/common_voice_7_0 with split test

python eval.py --model_id anuragshas/wav2vec2-large-xls-r-300m-pa-in --dataset mozilla-foundation/common_voice_7_0 --config pa-IN --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-pa-in"
sample_iter = iter(load_dataset("mozilla-foundation/common_voice_7_0", "pa-IN", 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
# => "ਉਨ੍ਹਾਂ ਨੇ ਸਾਰੇ ਤੇਅਰਵੇ ਵੱਖਰੀ ਕਿਸਮ ਦੇ ਕੀਤੇ ਹਨ"

Eval results on Common Voice 7 "test" (WER):

Without LM	With LM (run ./eval.py)
51.968	45.611