wav2vec2-large-xls-r-300m-as

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: 1.9068
• Wer: 0.6679

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
5.7027	21.05	400	3.4157	1.0
1.1638	42.1	800	1.3498	0.7461
0.2266	63.15	1200	1.6147	0.7273
0.1473	84.21	1600	1.6649	0.7108
0.1043	105.26	2000	1.7691	0.7090
0.0779	126.31	2400	1.8300	0.7009
0.0613	147.36	2800	1.8681	0.6916
0.0471	168.41	3200	1.8567	0.6875
0.0343	189.46	3600	1.9054	0.6840
0.0265	210.51	4000	1.9020	0.6786
0.0219	231.56	4400	1.9068	0.6679

Framework versions

• Transformers 4.16.0
• Pytorch 1.10.0+cu111
• Datasets 1.17.0
• 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-as --dataset mozilla-foundation/common_voice_7_0 --config as --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-as"
sample_iter = iter(load_dataset("mozilla-foundation/common_voice_7_0", "as", 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
# => "জাহাজত তো তিশকুৰলৈ যাব কিন্তু জহাজিটো আহিপনে"

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

Without LM	With LM (run ./eval.py)
67	56.995