This model is a fine-tuned version of openai/whisper-small on the Common Voice 11.0 dataset. It achieves 20.32 WER. Data augmentation can be implemented to further improve the model performance.

Intended uses & limitations

python

from datasets import load_dataset
from transformers import WhisperProcessor, WhisperForConditionalGeneration
from datasets import Audio

# load the dataset
test_dataset = load_dataset("mozilla-foundation/common_voice_11_0", "ar", split="test", use_auth_token=True, trust_remote_code=True)

# get the processor and model from mohammed/whisper-small-arabic-cv-11
processor = WhisperProcessor.from_pretrained("mohammed/whisper-small-arabic-cv-11")
model = WhisperForConditionalGeneration.from_pretrained("mohammed/whisper-small-arabic-cv-11")
model.config.forced_decoder_ids = None

# resample the audio files to 16000
test_dataset = test_dataset.cast_column("audio", Audio(sampling_rate=16000))

# get 10 exmaples of model transcription
for i in range(10):
  sample = test_dataset[i]["audio"]
  input_features = processor(sample["array"], sampling_rate=sample["sampling_rate"], return_tensors="pt").input_features 
  predicted_ids = model.generate(input_features)
  transcription = processor.batch_decode(predicted_ids, skip_special_tokens=False)
  transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
  print(f"{i} Reference Sentence: {test_dataset[i]['sentence']}")
  print(f"{i} Predicted Sentence: {transcription[0]}")

The output is:

0 Reference Sentence: زارني في أوائل الشهر بدري
0 Predicted Sentence: زارني في أوائل الشهر بدري 
1 Reference Sentence: إبنك بطل.
1 Predicted Sentence: ابنك بطل  
2 Reference Sentence: الواعظ الأمرد هذا الذي
2 Predicted Sentence: الوعز الأمرد هذا الذي 
3 Reference Sentence: سمح له هذا بالتخصص في البرونز الصغير، الذي يتم إنتاجه بشكل رئيسي ومربح للتصدير.
3 Predicted Sentence: صمح له هازب التخزوس في البرونز الصغير الذي زيت معنى به بشكل رئيسي من غربح للتصدير  
4 Reference Sentence: ألديك قلم ؟
4 Predicted Sentence: ألديك قلم  
5 Reference Sentence: يا نديمي قسم بي الى الصهباء
5 Predicted Sentence: يا نديمي قد سنبي إلى الصحباء  
6 Reference Sentence: إنك تكبر المشكلة.
6 Predicted Sentence: إنك تكبر المشكلة  
7 Reference Sentence: يرغب أن يلتقي بك.
7 Predicted Sentence: يرغب أن يلتقي بك  
8 Reference Sentence: إنهم لا يعرفون لماذا حتى.
8 Predicted Sentence: إنهم لا يعرفون لماذا حبت  
9 Reference Sentence: سيسعدني مساعدتك أي وقت تحب.
9 Predicted Sentence: سيسعد لمساعدتك أي وقت تحب

Training and evaluation data

This model is trained on the Common Voice 11.0 dataset.

Training procedure

The model is trained on 64 cores CPU, Nvidia 4070 Ti with 24 GB VRAM, and 100GB Disk space. The GPU utilization reached 100%. Please check the training hyperparameters below.

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 1e-05
train_batch_size: 2
eval_batch_size: 8
seed: 42
gradient_accumulation_steps: 16
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
training_steps: 5000
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Wer
0.721	0.2079	250	0.3651	29.8761
0.3044	0.4158	500	0.3308	27.6497
0.262	0.6237	750	0.3085	25.2769
0.2396	0.8316	1000	0.2863	24.5298
0.1998	1.0394	1250	0.2743	23.2776
0.134	1.2473	1500	0.2749	22.9829
0.1328	1.4552	1750	0.2662	22.3315
0.1314	1.6631	2000	0.2643	21.7402
0.1262	1.8710	2250	0.2598	21.8566
0.101	2.0789	2500	0.2608	21.4248
0.0653	2.2868	2750	0.2682	20.9912
0.062	2.4947	3000	0.2638	21.0137
0.0627	2.7026	3250	0.2636	20.5369
0.0603	2.9105	3500	0.2602	20.4580
0.0456	3.1183	3750	0.2748	20.9555
0.0324	3.3262	4000	0.2702	20.4918
0.0318	3.5341	4250	0.2739	20.4355
0.0296	3.7420	4500	0.2735	20.4374
0.0291	3.9499	4750	0.2725	20.3717
0.022	4.1578	5000	0.2758	20.3229

Framework versions

Transformers 4.42.4
Pytorch 2.3.1+cu121
Datasets 2.20.0
Tokenizers 0.19.1