wav2vec2-large-xls-r-300m-urdu Open Source Model - Free Deployment for Precise Urdu Speech Recognition

Wav2vec2 Large Xls R 300m Urdu

Developed by omar47

This model is a fine-tuned version based on facebook/wav2vec2-xls-r-300m, specifically optimized for Urdu speech recognition tasks.

Speech Recognition

Transformers

Open Source License:Apache-2.0 #Urdu speech recognition #Low word error rate #Large model fine-tuning

Downloads 27

Release Time : 4/29/2022

Model Overview

This is an automatic speech recognition (ASR) model for Urdu, fine-tuned on the wav2vec2-xls-r-300m architecture, achieving a 17.02% word error rate (WER) on the evaluation set.

Model Features

Urdu optimization

Fine-tuned specifically for Urdu speech characteristics to improve recognition accuracy

Low word error rate

Achieved a 17.02% word error rate (WER) on the evaluation set

Fine-tuned from a large model

Fine-tuned from the facebook/wav2vec2-xls-r-300m large model, inheriting its powerful speech feature extraction capabilities

Model Capabilities

Urdu speech recognition

Audio to text

Speech transcription

Use Cases

Speech transcription

Urdu meeting minutes

Automatically transcribe Urdu meeting recordings into text records

Approximately 83% accuracy

Voice assistant

Speech recognition component for Urdu voice assistant applications

Education

Language learning applications

Help learners practice Urdu pronunciation and provide feedback

🚀 wav2vec2-large-xls-r-300m-urdu

This model is a fine - tuned version of facebook/wav2vec2-xls-r-300m. It is designed for specific Urdu - related speech tasks and offers improved performance on the evaluation set.

🚀 Quick Start

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

Loss: 0.5285
Wer: 0.1702

🔧 Technical Details

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0003
train_batch_size: 16
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 2
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: 500
num_epochs: 35
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
16.9618	0.74	32	15.0745	1.0
9.1928	1.49	64	5.9361	1.0
4.9307	2.23	96	4.2924	1.0
3.8917	2.98	128	3.5873	1.0
3.3867	3.72	160	3.2594	1.0
3.2107	4.47	192	3.1718	1.0
3.1395	5.21	224	3.1281	1.0
3.115	5.95	256	3.1238	1.0
3.0801	6.7	288	3.0674	1.0
2.9725	7.44	320	2.8277	1.0
2.4159	8.19	352	1.7186	0.9036
1.3377	8.93	384	1.0271	0.6433
0.8591	9.67	416	0.8087	0.5441
0.726	10.42	448	0.7263	0.4634
0.6242	11.16	480	0.6783	0.4156
0.5417	11.91	512	0.6611	0.4305
0.4784	12.65	544	0.6300	0.3926
0.4198	13.4	576	0.5646	0.3499
0.3798	14.14	608	0.5919	0.3229
0.3356	14.88	640	0.5715	0.3369
0.2954	15.63	672	0.5325	0.2728
0.264	16.37	704	0.5535	0.2689
0.2535	17.12	736	0.5467	0.2366
0.2277	17.86	768	0.5219	0.2345
0.2141	18.6	800	0.5314	0.2487
0.2036	19.35	832	0.5382	0.2236
0.2021	20.09	864	0.5038	0.1922
0.1676	20.84	896	0.5238	0.2033
0.1544	21.58	928	0.5069	0.1866
0.1512	22.33	960	0.5045	0.1965
0.1512	23.07	992	0.5167	0.1862
0.1399	23.81	1024	0.5236	0.1840
0.1291	24.56	1056	0.5234	0.1957
0.1274	25.3	1088	0.5348	0.1943
0.127	26.05	1120	0.4978	0.1719
0.1105	26.79	1152	0.5067	0.1767
0.1069	27.53	1184	0.5150	0.1758
0.1058	28.28	1216	0.5218	0.1844
0.0999	29.02	1248	0.5375	0.1852
0.0964	29.77	1280	0.5373	0.1843
0.0971	30.51	1312	0.5190	0.1776
0.0906	31.26	1344	0.5217	0.1747
0.0909	32.0	1376	0.5204	0.1778
0.0784	32.74	1408	0.5336	0.1756
0.0823	33.49	1440	0.5281	0.1699
0.0834	34.23	1472	0.5292	0.1700
0.0827	34.98	1504	0.5285	0.1702

Framework versions

Transformers 4.18.0
Pytorch 1.11.0+cu113
Datasets 2.1.0
Tokenizers 0.12.1

📄 License

This project is licensed under the Apache - 2.0 license.

Featured Recommended AI Models

Empowering the Future, Your AI Solution Knowledge Base

English 简体中文繁體中文にほんご