XLS-R-300M - Slovak

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

Loss: 0.3067
Wer: 0.2678

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: 1500
num_epochs: 60.0
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
5.175	2.41	400	4.6909	1.0
3.3785	4.82	800	3.3080	1.0
2.6964	7.23	1200	2.0651	1.1055
1.3008	9.64	1600	0.5845	0.6207
1.1185	12.05	2000	0.4195	0.4193
1.0252	14.46	2400	0.3824	0.3570
0.935	16.87	2800	0.3693	0.3462
0.8818	19.28	3200	0.3587	0.3318
0.8534	21.69	3600	0.3420	0.3180
0.8137	24.1	4000	0.3426	0.3130
0.7968	26.51	4400	0.3349	0.3102
0.7558	28.92	4800	0.3216	0.3019
0.7313	31.33	5200	0.3451	0.3060
0.7358	33.73	5600	0.3272	0.2967
0.718	36.14	6000	0.3315	0.2882
0.6991	38.55	6400	0.3299	0.2830
0.6529	40.96	6800	0.3140	0.2836
0.6225	43.37	7200	0.3128	0.2751
0.633	45.78	7600	0.3211	0.2774
0.5876	48.19	8000	0.3162	0.2764
0.588	50.6	8400	0.3082	0.2722
0.5915	53.01	8800	0.3120	0.2681
0.5798	55.42	9200	0.3133	0.2709
0.5736	57.83	9600	0.3086	0.2676

Framework versions

Transformers 4.17.0.dev0
Pytorch 1.10.2+cu102
Datasets 1.18.4.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-sk-cv8-with-lm --dataset mozilla-foundation/common_voice_8_0 --config sk --split test

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

python eval.py --model_id anuragshas/wav2vec2-xls-r-300m-sk-cv8-with-lm --dataset speech-recognition-community-v2/dev_data --config sk --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-sk-cv8-with-lm"
sample_iter = iter(load_dataset("mozilla-foundation/common_voice_8_0", "sk", 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
# => ""

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

Without LM	With LM (run `./eval.py`)
26.707	18.609