XLS-R-300M - Bulgarian

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

Loss: 0.2473
Wer: 0.3002

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: 50.0
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
3.1589	3.48	400	3.0830	1.0
2.8921	6.96	800	2.6605	0.9982
1.3049	10.43	1200	0.5069	0.5707
1.1349	13.91	1600	0.4159	0.5041
1.0686	17.39	2000	0.3815	0.4746
0.999	20.87	2400	0.3541	0.4343
0.945	24.35	2800	0.3266	0.4132
0.9058	27.83	3200	0.2969	0.3771
0.8672	31.3	3600	0.2802	0.3553
0.8313	34.78	4000	0.2662	0.3380
0.8068	38.26	4400	0.2528	0.3181
0.7796	41.74	4800	0.2537	0.3073
0.7621	45.22	5200	0.2503	0.3036
0.7611	48.7	5600	0.2477	0.2991

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-large-xls-r-300m-bg --dataset mozilla-foundation/common_voice_8_0 --config bg --split test

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

python eval.py --model_id anuragshas/wav2vec2-large-xls-r-300m-bg --dataset speech-recognition-community-v2/dev_data --config bg --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-large-xls-r-300m-bg"
sample_iter = iter(load_dataset("mozilla-foundation/common_voice_8_0", "bg", 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
# => "и надутият му ката блоонкурем взе да се събира"

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

Without LM	With LM (run `./eval.py`)
30.07	21.195