whisper-large-v2-arabic-5k-steps

This model is a fine-tuned version of openai/whisper-large-v2 on the Arabic CommonVoice dataset (v11). It achieves the following results on the evaluation set:

• Loss: 0.3434
• Wer: 0.4239

Model description

This model is finetuned for 5000 steps for research purposes which means that the transcriptions might not be that satisfactory for users.

Training and evaluation data

• Training Data: CommonVoice (v11) train split
• Validation Data: CommonVoice (v11) Validation split
• Test Data: CommonVoice (v11) Test split

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 50
• eval_batch_size: 16
• seed: 42
• 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
0.1638	1.78	1000	0.2295	0.4410
0.0587	3.57	2000	0.2337	0.4272
0.0125	5.35	3000	0.2745	0.4208
0.004	7.13	4000	0.3124	0.4252
0.0016	8.91	5000	0.3434	0.4239

Transcription:

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

# device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# load the model
processor = WhisperProcessor.from_pretrained("clu-ling/whisper-large-v2-arabic-5k-steps")
model = WhisperForConditionalGeneration.from_pretrained("clu-ling/whisper-large-v2-arabic-5k-steps").to(device)
forced_decoder_ids = processor.get_decoder_prompt_ids(language="ar", task="transcribe")

# load the dataset
commonvoice_eval = load_dataset("mozilla-foundation/common_voice_11_0", "ar", split="validation", streaming=True)
commonvoice_eval = commonvoice_eval.cast_column("audio", Audio(sampling_rate=16000))
sample = next(iter(commonvoice_eval))["audio"]

# features and generate token ids
input_features = processor(sample["array"], sampling_rate=sample["sampling_rate"], return_tensors="pt").input_features
predicted_ids = model.generate(input_features.to(device), forced_decoder_ids=forced_decoder_ids)

# decode
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]

print("Transcription:", transcription)
Transcription: عمي هو أخو أبي.

Evaluation:

Evaluates this model on mozilla-foundation/common_voice_11_0 test split.

import pyarabic.araby as araby
from transformers.models.whisper.english_normalizer import BasicTextNormalizer
from datasets import load_dataset, Audio
import evaluate
import torch
import re
from transformers import WhisperProcessor, WhisperForConditionalGeneration

# device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# metric
wer_metric = evaluate.load("wer")

# model
processor = WhisperProcessor.from_pretrained("clu-ling/whisper-large-v2-arabic-5k-steps")
model = WhisperForConditionalGeneration.from_pretrained("clu-ling/whisper-large-v2-arabic-5k-steps")

# dataset
dataset = load_dataset("mozilla-foundation/common_voice_11_0", "ar", split="test", ) #cache_dir=args.cache_dir
dataset = dataset.cast_column("audio", Audio(sampling_rate=16000))

#for debuggings: it gets two examples
#dataset = dataset.shard(num_shards=10000, index=0)
#print(dataset)

def clean_text(text):
  """Normalizes TRANSCRIPT"""
  text = re.sub(r'[\,\?\.\!\-\;\:\"\“\%\٪\‘\”\�\«\»\،\.\:\؟\؛\*\>\<]', '', text) + " " # special characters
  text = re.sub(r'http\S+', '', text) + " " # links
  text = re.sub(r'[\[\]\(\)\-\/\{\}]', '', text) + " " # brackets
  text = re.sub(r'\s+', ' ', text) + " " # extra white space
  text = araby.strip_diacritics(text) # remove diacrirics
  return text.strip()
    
def normalize(batch):
  """Normalizes GOLD"""
  #batch["gold_text"] = whisper_norm(batch['sentence'])
  batch["gold_text"] = clean_text(batch['sentence'])
  return batch

def map_wer(batch):
  model.to(device)
  forced_decoder_ids = processor.get_decoder_prompt_ids(language = "ar", task = "transcribe")
  inputs = processor(batch["audio"]["array"], sampling_rate=batch["audio"]["sampling_rate"], return_tensors="pt").input_features
  with torch.no_grad():
    generated_ids = model.generate(inputs=inputs.to(device), forced_decoder_ids=forced_decoder_ids)
    transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
  batch["predicted_text"] = clean_text(transcription)
  return batch

# process GOLD text
processed_dataset = dataset.map(normalize)
# get predictions
predicted = processed_dataset.map(map_wer)

# word error rate
wer = wer_metric.compute(references=predicted['gold_text'], predictions=predicted['predicted_text'])
wer = round(100 * wer, 2)
print("WER:", wer)

Framework versions

• Transformers 4.26.0.dev0
• Pytorch 1.13.1
• Datasets 2.8.1.dev0
• Tokenizers 0.13.2