prepare_nemo_manifest.py

from datasets import Dataset, DatasetDict, IterableDataset, IterableDatasetDict, concatenate_datasets, load_dataset
import re
import soundfile as sf

TDataset = DatasetDict | Dataset | IterableDatasetDict | IterableDataset

def remove_special_tags(dataset: TDataset, label_column_name: str = 'words', **kwargs) -> TDataset:
    special_tag_pattern = r'(\[\w+( [0-9])?:?(\w+)?\] )'
    def remove_special_tags(batch):
        batch[label_column_name] = re.sub(special_tag_pattern, '', batch[label_column_name]).lower()
        return batch
    
    dataset = dataset.map(remove_special_tags, **kwargs)
    
    return dataset

def remove_bad_files(dataset: TDataset, **kwargs) -> TDataset:
    def is_file_good(batch):
        info = sf.info(batch['path'])
        if info.channels == 0:
            print(f"File {batch['path']} has no channels. Removing it from the dataset.")
            return False
        if info.duration == 0:
            print(f"File {batch['path']} has duration 0. Removing it from the dataset.")
            return False

        return True
    
    dataset = dataset.filter(is_file_good, **kwargs)
    
    return dataset

if __name__ == "__main__":
    print("loading dataset...")
    dataset = load_dataset('./transcribed_calls.py', trust_remote_code=True)
    print('removing bad files...')
    dataset = remove_bad_files(dataset)
    print("removing special tags...")
    dataset = remove_special_tags(dataset)
    print("concatenating train and test datasets...")
    dataset = concatenate_datasets([dataset['train'], dataset['test']])

    print("creating nemo_manifest.json...")
    with open('nemo_manifest.json', 'w') as f:
        for example in dataset:
            f.write(f'{{"audio_filepath": "{example["path"]}", "text": "{example["words"]}"}}\n')

    print('nemo_manifest.json created')