XLS-R-1B - Hindi

This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - HI dataset. It achieves the following results on the evaluation set:

Loss: 0.6921
Wer: 0.3547

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: 5e-05
train_batch_size: 8
eval_batch_size: 16
seed: 42
gradient_accumulation_steps: 4
total_train_batch_size: 32
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 1500
num_epochs: 50.0
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
2.0674	2.07	400	1.3411	0.8835
1.324	4.15	800	0.9311	0.7142
1.2023	6.22	1200	0.8060	0.6170
1.1573	8.29	1600	0.7415	0.4972
1.1117	10.36	2000	0.7248	0.4588
1.0672	12.44	2400	0.6729	0.4350
1.0336	14.51	2800	0.7117	0.4346
1.0025	16.58	3200	0.7019	0.4272
0.9578	18.65	3600	0.6792	0.4118
0.9272	20.73	4000	0.6863	0.4156
0.9321	22.8	4400	0.6535	0.3972
0.8802	24.87	4800	0.6766	0.3906
0.844	26.94	5200	0.6782	0.3949
0.8387	29.02	5600	0.6916	0.3921
0.8042	31.09	6000	0.6806	0.3797
0.793	33.16	6400	0.7120	0.3831
0.7567	35.23	6800	0.6862	0.3808
0.7463	37.31	7200	0.6893	0.3709
0.7053	39.38	7600	0.7096	0.3701
0.6906	41.45	8000	0.6921	0.3676
0.6891	43.52	8400	0.7167	0.3663
0.658	45.6	8800	0.6833	0.3580
0.6576	47.67	9200	0.6914	0.3569
0.6358	49.74	9600	0.6922	0.3551

Framework versions

Transformers 4.16.0.dev0
Pytorch 1.10.1+cu102
Datasets 1.17.1.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-1b-hi-with-lm --dataset mozilla-foundation/common_voice_8_0 --config hi --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-1b-hi-with-lm"
sample_iter = iter(load_dataset("mozilla-foundation/common_voice_8_0", "hi", 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
# => "तुम्हारे पास तीन महीने बचे हैं"

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

Without LM	With LM (run `./eval.py`)
26.209	15.899