This model is a fine-tuned version of facebook/wav2vec2-xls-r-1b on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - HI dataset. It achieves the following results on the evaluation set:

Loss: 0.5878
Wer: 0.3419

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: 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: 2000
num_epochs: 100.0
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
1.9859	2.72	400	1.1663	0.7948
1.2969	5.44	800	0.7725	0.6562
1.1954	8.16	1200	0.5940	0.4904
1.164	10.88	1600	0.5338	0.4316
1.1464	13.6	2000	0.5432	0.4226
1.1553	16.33	2400	0.5471	0.4260
1.0985	19.05	2800	0.5290	0.4076
1.0421	21.77	3200	0.5672	0.4181
0.9831	24.49	3600	0.5741	0.4141
0.9827	27.21	4000	0.5754	0.4179
0.9669	29.93	4400	0.5310	0.3889
0.9496	32.65	4800	0.5649	0.4062
0.9112	35.37	5200	0.5738	0.3926
0.8838	38.1	5600	0.5232	0.3768
0.8666	40.81	6000	0.5510	0.3852
0.8366	43.54	6400	0.5436	0.3837
0.7957	46.26	6800	0.5337	0.3775
0.7834	48.98	7200	0.5611	0.3844
0.7685	51.7	7600	0.5710	0.4008
0.7431	54.42	8000	0.5636	0.3726
0.7353	57.14	8400	0.5937	0.3836
0.7001	59.86	8800	0.5815	0.3858
0.6799	62.58	9200	0.5862	0.3696
0.6459	65.31	9600	0.6181	0.3762
0.6121	68.03	10000	0.5637	0.3590
0.5942	70.75	10400	0.6374	0.3882
0.5769	73.47	10800	0.6015	0.3640
0.5689	76.19	11200	0.5669	0.3508
0.5461	78.91	11600	0.5967	0.3621
0.5286	81.63	12000	0.5840	0.3605
0.5057	84.35	12400	0.5848	0.3489
0.482	87.07	12800	0.5860	0.3488
0.4655	89.79	13200	0.5780	0.3453
0.4523	92.52	13600	0.6150	0.3532
0.4422	95.24	14000	0.5930	0.3452
0.4436	97.96	14400	0.5867	0.3428

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_7_0 with split test

python eval.py --model_id anuragshas/wav2vec2-xls-r-1b-hi --dataset mozilla-foundation/common_voice_7_0 --config hi --split test

Inference With LM

python

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"
sample_iter = iter(load_dataset("mozilla-foundation/common_voice_7_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 7 "test" (WER):

Without LM	With LM (run `./eval.py`)
28.942	18.504