anuragshas/wav2vec2-xls-r-300m-mt-cv8-with-lm

XLS-R-300M - Maltese

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - MT dataset. It achieves the following results on the evaluation set:

• Loss: 0.1895
• Wer: 0.1984

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: 7.5e-05
• train_batch_size: 32
• eval_batch_size: 16
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 1000
• num_epochs: 60.0
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.4219	3.6	400	3.3127	1.0
3.0399	7.21	800	3.0330	1.0
1.5756	10.81	1200	0.6108	0.5724
1.0995	14.41	1600	0.3091	0.3154
0.9639	18.02	2000	0.2596	0.2841
0.9032	21.62	2400	0.2270	0.2514
0.8145	25.23	2800	0.2172	0.2483
0.7845	28.83	3200	0.2084	0.2333
0.7694	32.43	3600	0.1974	0.2234
0.7333	36.04	4000	0.2020	0.2185
0.693	39.64	4400	0.1947	0.2148
0.6802	43.24	4800	0.1960	0.2102
0.667	46.85	5200	0.1904	0.2072
0.6486	50.45	5600	0.1881	0.2009
0.6339	54.05	6000	0.1877	0.1989
0.6254	57.66	6400	0.1893	0.2003

Framework versions

• Transformers 4.17.0.dev0
• Pytorch 1.10.2+cu102
• Datasets 1.18.2.dev0
• Tokenizers 0.11.0

Evaluation Commands

1. To evaluate on mozilla-foundation/common_voice_8_0 with split test

python eval.py --model_id anuragshas/wav2vec2-xls-r-300m-mt-cv8-with-lm --dataset mozilla-foundation/common_voice_8_0 --config mt --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-xls-r-300m-mt-cv8-with-lm"
sample_iter = iter(load_dataset("mozilla-foundation/common_voice_8_0", "mt", 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
# => "għadu jilagħbu ċirku tant bilfondi"

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

Without LM	With LM (run ./eval.py)
19.853	15.967