Aradia-ctc-distilhubert-ft Open Source Model - Supports Automatic Arabic Speech Recognition

Home

Aradia Ctc Distilhubert Ft

Developed by abdusah

An automatic speech recognition (ASR) model fine-tuned on Arabic speech datasets based on distilhubert

Speech Recognition

Transformers

Open Source License:Apache-2.0 #Arabic speech recognition #Lightweight HuBERT #Low-resource optimization

Downloads 16

Release Time : 4/6/2022

Model Overview

This model is a lightweight speech recognition model based on the distilhubert architecture, specifically optimized for Arabic speech recognition tasks.

Model Features

Lightweight architecture

Lightweight design based on DistilHuBERT, suitable for resource-constrained environments

Arabic optimization

Specifically fine-tuned for Arabic speech characteristics

Efficient training

Uses mixed-precision training and gradient accumulation techniques to improve training efficiency

Model Capabilities

Arabic speech-to-text

Continuous speech recognition

Large-scale speech data processing

Use Cases

Speech transcription

Arabic speech transcription

Convert Arabic speech content into text

Word Error Rate (WER) 0.8908

Voice assistants

Arabic voice command recognition

Used for front-end speech recognition in Arabic voice assistants

🚀 aradia-ctc-distilhubert-ft

This model is a fine - tuned version of [ntu - spml/distilhubert](https://huggingface.co/ntu - spml/distilhubert) on the ABDUSAHMBZUAI/ARABIC_SPEECH_MASSIVE_SM - NA dataset. It offers enhanced performance in automatic speech recognition tasks. On the evaluation set, it achieves the following results:

Loss: 2.7114
Wer: 0.8908

📚 Documentation

Model description

This model is a fine - tuned variant of [ntu - spml/distilhubert](https://huggingface.co/ntu - spml/distilhubert) on the ABDUSAHMBZUAI/ARABIC_SPEECH_MASSIVE_SM - NA dataset.

Intended uses & limitations

The model is mainly intended for automatic speech recognition tasks. However, more information about its limitations needs to be provided.

Training and evaluation data

Details about the training and evaluation data are yet to be provided.

🔧 Technical Details

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0003
train_batch_size: 32
eval_batch_size: 32
seed: 42
gradient_accumulation_steps: 2
total_train_batch_size: 64
optimizer: Adam with betas=(0.9,0.999) and epsilon = 1e - 08
lr_scheduler_type: linear
lr_scheduler_warmup_steps: 500
num_epochs: 30.0
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
No log	0.43	100	4.4129	1.0
No log	0.87	200	3.5927	1.0
No log	1.3	300	3.3780	1.0
No log	1.74	400	3.0830	1.0
5.3551	2.17	500	2.6278	0.9999
5.3551	2.61	600	1.8359	1.0000
5.3551	3.04	700	1.7878	0.9914
5.3551	3.48	800	1.5219	0.9875
5.3551	3.91	900	1.4348	0.9879
1.7199	4.35	1000	1.4354	0.9644
1.7199	4.78	1100	1.5210	0.9519
1.7199	5.22	1200	1.3607	0.9475
1.7199	5.65	1300	1.3839	0.9343
1.7199	6.09	1400	1.2806	0.8944
1.2342	6.52	1500	1.3036	0.9011
1.2342	6.95	1600	1.3704	0.9072
1.2342	7.39	1700	1.2981	0.8891
1.2342	7.82	1800	1.2786	0.8733
1.2342	8.26	1900	1.2897	0.8867
0.9831	8.69	2000	1.4436	0.8780
0.9831	9.13	2100	1.3680	0.8873
0.9831	9.56	2200	1.3471	0.8692
0.9831	10.0	2300	1.3725	0.8729
0.9831	10.43	2400	1.4439	0.8771
0.8071	10.87	2500	1.5114	0.8928
0.8071	11.3	2600	1.6156	0.8958
0.8071	11.74	2700	1.4381	0.8749
0.8071	12.17	2800	1.5088	0.8717
0.8071	12.61	2900	1.5486	0.8813
0.6321	13.04	3000	1.4536	0.8884
0.6321	13.48	3100	1.4679	0.8947
0.6321	13.91	3200	1.5628	0.9117
0.6321	14.35	3300	1.5831	0.8716
0.6321	14.78	3400	1.6733	0.8702
0.4998	15.22	3500	1.8225	0.8665
0.4998	15.65	3600	1.8558	0.8732
0.4998	16.09	3700	1.7513	0.8766
0.4998	16.52	3800	1.8562	0.8753
0.4998	16.95	3900	1.9018	0.8704
0.4421	17.39	4000	1.9341	0.8789
0.4421	17.82	4100	1.9582	0.8781
0.4421	18.26	4200	1.8863	0.8821
0.4421	18.69	4300	1.9366	0.8847
0.4421	19.13	4400	2.1902	0.8721
0.3712	19.56	4500	2.1641	0.8670
0.3712	20.0	4600	2.1639	0.8776
0.3712	20.43	4700	2.2695	0.9030
0.3712	20.87	4800	2.1909	0.8937
0.3712	21.3	4900	2.1606	0.8959
0.3067	21.74	5000	2.1756	0.8943
0.3067	22.17	5100	2.4092	0.8773
0.3067	22.61	5200	2.4991	0.8721
0.3067	23.04	5300	2.3340	0.8910
0.3067	23.48	5400	2.3567	0.8946
0.2764	23.91	5500	2.3215	0.8897
0.2764	24.35	5600	2.4824	0.9002
0.2764	24.78	5700	2.4585	0.8963
0.2764	25.22	5800	2.5804	0.8879
0.2764	25.65	5900	2.5814	0.8903
0.2593	26.09	6000	2.5374	0.8868
0.2593	26.52	6100	2.5346	0.8922
0.2593	26.95	6200	2.5465	0.8873
0.2593	27.39	6300	2.6002	0.8919
0.2593	27.82	6400	2.6102	0.8928
0.227	28.26	6500	2.6925	0.8914
0.227	28.69	6600	2.6981	0.8913
0.227	29.13	6700	2.6872	0.8891
0.227	29.56	6800	2.7015	0.8897
0.227	30.0	6900	2.7114	0.8908

Framework versions

Transformers 4.18.0.dev0
Pytorch 1.10.2+cu113
Datasets 1.18.4
Tokenizers 0.11.6

📄 License

This model is released 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