infer-wer_fleurs.py

from datasets import load_dataset
from transformers.pipelines.pt_utils import KeyDataset
from transformers import pipeline as pipesetup
import tqdm
import jiwer

# Use a streaming dataset
fleurs = load_dataset("google/fleurs", "bn_in", split="train", streaming=True)

# Get an iterator for the dataset
iterator = iter(fleurs)

pred, gt = [], []

pipeline = pipesetup(model="ai4bharat/indicwav2vec_v1_bengali")

for i in tqdm.tqdm(range(1000)):  # Adjust the range according to your requirement
    # Fetch one item from the dataset
    data = next(iterator)
    #breakpoint()
    inp, out = data["raw_transcription"], pipeline(data["audio"]['array'], max_new_tokens=448)
    gt.append(inp)
    pred.append(out['text'])
    if (i+1) % 10 == 0:
        print(f'WER after {i+1} samples: {jiwer.wer(gt,pred)}')

# WER computation after all the samples have been processed
final_wer = jiwer.wer(gt, pred)
print(f'Final WER: {final_wer}')