XLS-R-300M - Slovenian

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

Loss: 0.2578
Wer: 0.2273

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
3.1829	4.88	400	3.1228	1.0
2.8675	9.76	800	2.8616	0.9993
1.583	14.63	1200	0.6392	0.6239
1.1959	19.51	1600	0.3602	0.3651
1.0276	24.39	2000	0.3021	0.2981
0.9671	29.27	2400	0.2872	0.2739
0.873	34.15	2800	0.2593	0.2459
0.8513	39.02	3200	0.2617	0.2473
0.8132	43.9	3600	0.2548	0.2426
0.7935	48.78	4000	0.2637	0.2353
0.7565	53.66	4400	0.2629	0.2322
0.7359	58.54	4800	0.2579	0.2253

Framework versions

Transformers 4.17.0.dev0
Pytorch 1.10.2+cu102
Datasets 1.18.2.dev0
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-xls-r-300m-sl-cv8-with-lm --dataset mozilla-foundation/common_voice_8_0 --config sl --split test

To evaluate on speech-recognition-community-v2/dev_data

python eval.py --model_id anuragshas/wav2vec2-xls-r-300m-sl-cv8-with-lm --dataset speech-recognition-community-v2/dev_data --config sl --split validation --chunk_length_s 5.0 --stride_length_s 1.0

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-sl-cv8-with-lm"
sample_iter = iter(load_dataset("mozilla-foundation/common_voice_8_0", "sl", 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
# => "zmago je divje od letel s helikopterjem visoko vzrak"

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

Without LM	With LM (run `./eval.py`)
19.938	12.736