wav2vec2-large-xls-r-300m-mr

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: 0.5479
Wer: 0.5740

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
3.7378	18.18	400	3.5047	1.0
3.1707	36.36	800	2.6166	0.9912
1.4942	54.55	1200	0.5778	0.6927
1.2058	72.73	1600	0.5168	0.6362
1.0558	90.91	2000	0.5105	0.6069
0.9488	109.09	2400	0.5151	0.6089
0.8588	127.27	2800	0.5157	0.5989
0.7991	145.45	3200	0.5179	0.5740
0.7545	163.64	3600	0.5348	0.5740
0.7144	181.82	4000	0.5518	0.5724
0.7041	200.0	4400	0.5479	0.5740

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-mr --dataset mozilla-foundation/common_voice_8_0 --config mr --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-mr"
sample_iter = iter(load_dataset("mozilla-foundation/common_voice_8_0", "mr", 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
# => "या पानास लेखाचे स्वरूप यायला हावे"

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

Without LM	With LM (run `./eval.py`)
49.177	32.811