wav2vec2-large-xls-r-300m-or

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: 1.6618
• Wer: 0.5166

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: 240

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
6.0493	23.53	400	2.9728	1.0
0.5306	47.06	800	1.2895	0.6138
0.1253	70.59	1200	1.6854	0.5703
0.0763	94.12	1600	1.9433	0.5870
0.0552	117.65	2000	1.4393	0.5575
0.0382	141.18	2400	1.4665	0.5537
0.0286	164.71	2800	1.5441	0.5320
0.0212	188.24	3200	1.6502	0.5115
0.0168	211.76	3600	1.6411	0.5332
0.0129	235.29	4000	1.6618	0.5166

Framework versions

• Transformers 4.16.0
• 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-or --dataset mozilla-foundation/common_voice_7_0 --config or --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-or"
sample_iter = iter(load_dataset("mozilla-foundation/common_voice_7_0", "or", 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
# => "ପରରାଏ ବାଲା ଗସ୍ତି ଫାଣ୍ଡି ଗୋପାଳ ପରଠାରୁ ଦେଢ଼କଶ ଦୂର"

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

Without LM	With LM (run ./eval.py)
51.92	47.186