Inconsistent wer on malay language compared to result published

[josephwong14wkh]

I am currently evaluating the performance of transcription on malay language (FLEUR testing dataset). However, my result is not the same as the published result. Here is the wer i got:

model size: wer:
medium 13.218%
large-v2 9.497%

Here is the published wer:
model size: wer:
medium 12.2%
large-v2 8.7%

Here is my script:
from jiwer import wer
from whisper.normalizers import BasicTextNormalizer

def cal_wer(model_size):
df = pd.read_csv(gtfile, header=None, sep="\t")
model = whisper.load_model(model_size)
gt_texts = list()
pred_texts = list()

for i in range(len(df)):
    # inference
    filename = df.iloc[i, 1]  # second column is filename
    _audio = whisper.load_audio(f"{audiodir}/{filename}")
    _audio = whisper.pad_or_trim(_audio)
    mel = whisper.log_mel_spectrogram(_audio).to(model.device)
    options = whisper.DecodingOptions(language='ms', fp16=False)
    result = whisper.decode(model, mel, options)
    pred_text = text_normalizer(result.text)
    gt_text = text_normalizer(df.iloc[i, 3]) # forth column is ground truth text (pre-normalized)
    
    print(f"gt_texts: {gt_text}")
    print(f"pred_texts: {pred_text}")
    gt_texts.append(gt_text)
    pred_texts.append(pred_text)
    
error = wer(gt_texts, pred_texts) 

I use https://github.com/jitsi/jiwer to evaluate the wer between ground truth text and STT text and use text normalizer provided in whisper/normalizers/basic.py. May i know is there any configuration (properly on DecodingOption?) that i need to set so that i can replicate the published result?