README.md

metadata

language:
  - ar
  - multilingual
license: apache-2.0
tags:
  - automatic-speech-recognition
  - hf-asr-leaderboard
  - whisper-event
  - generated_from_trainer
  - Arabic
  - multilingual
  - STT
datasets:
  - mozilla-foundation/common_voice_12_0
metrics:
  - wer
model-index:
  - name: Kalemat-Tech Arabic Speech Recognition Model (STT)
    results:
      - task:
          type: automatic-speech-recognition
          name: Speech Recognition
        dataset:
          type: mozilla-foundation/common_voice_12_0
          name: mozilla-foundation/common_voice_12_0
          config: ar
          split: test
          args: ar
        metrics:
          - type: wer
            value: 58.5848
            name: wer

Kalemat-Tech Arabic Speech Recognition Model (STT) - Mohamed Salama

نموذج كلماتك للتعرف على الأصوات العربية الفصحى و تحويلها إلى نصوص

KalemaTech-Arabic-STT-ASR-based-on-Whisper-Small

This model is a fine-tuned version of openai/whisper-small on Common_Voice_Arabic_12.0_Augmented. It achieves the following results on the evaluation set:

• Loss: 0.5362
• Wer: 58.5848

Example of usage:

from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq

processor = AutoProcessor.from_pretrained("Salama1429/KalemaTech-Arabic-STT-ASR-based-on-Whisper-Small")

model = AutoModelForSpeechSeq2Seq.from_pretrained("Salama1429/KalemaTech-Arabic-STT-ASR-based-on-Whisper-Small")

Intended uses & limitations

Automatic Speech Recognition

Training and evaluation data

Common_Voice_Arabic_12.0 and I made some augmentations to it as follows:
- 25% of the data TimeMasking
- 25% of the data SpecAugmentation
- 25% of the data WavAugmentation (AddGaussianNoise)
- The final dataset is the original common voice plus the augmented files

Training procedure

Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 1e-05
- train_batch_size: 64
- eval_batch_size: 8
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- num_epochs: 25
- mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.2728	1.01	1000	0.3063	60.4733
0.1442	2.01	2000	0.2878	55.6935
0.0648	3.02	3000	0.3009	59.2568
0.0318	4.03	4000	0.3278	59.2993
0.0148	5.04	5000	0.3539	61.0364
0.0088	6.04	6000	0.3714	56.9154
0.0061	7.05	7000	0.3920	57.5515
0.0041	8.06	8000	0.4149	61.6328
0.0033	9.06	9000	0.4217	58.0310
0.0033	10.07	10000	0.4376	59.9594
0.0021	11.08	11000	0.4485	56.7812
0.0015	12.08	12000	0.4577	57.6936
0.0013	13.09	13000	0.4671	60.6606
0.0011	14.1	14000	0.4686	59.8159
0.0008	15.11	15000	0.4856	60.7111
0.0011	16.11	16000	0.4851	59.5198
0.0005	17.12	17000	0.4936	59.2608
0.0004	18.13	18000	0.4995	57.9619
0.0003	19.13	19000	0.5085	58.3630
0.0002	20.14	20000	0.5155	58.0987
0.0001	21.15	21000	0.5251	58.8504
0.0001	22.16	22000	0.5268	58.4228
0.0001	23.16	23000	0.5317	59.0881
0.0001	24.17	24000	0.5362	58.5848

Framework versions

• Transformers 4.25.1
• Pytorch 1.13.1+cu117
• Datasets 2.8.0
• Tokenizers 0.13.2