Unsupervised_Metrics_CLD_Sy/code/make_noiser.py

"""This module inmplements some methods to artificialy noise the data
"""

from typing import List, Dict
from random import choices, sample, shuffle
import copy
import random
random.seed(80)

class Noise :
    """
    This class simulate noise in the data. Crucially, noise can be made on three points :\
    (1) The noise of phonemes order of a given sequence by making the order of the sequence more aribitrary,\
    (2) Replacement of some phonemes of a given sequence by arbitrary sampled phonemes from a vocabulary and\
    (3) By arbitrary interverting some sequences of two different speakers.

    Atributes
    ---------
    - phonemes_order_noise :
        Parameter for controling the degree of noise at the level of phonemes order. See the point 1 mentioned above.
    - speakers_noise :
        Parameters for controling the degree of noise at the level of speakers. See the point 3 mentioned above.
    - phonemes_noise :
        Parameter for controling the degree of noise at the level of phonemes. See the point 2 mentioned above.
    """

    def __init__(self,
                    most_probable_phonemes: list,
                    phonemes_order_noise=0.3,
                    speakers_noise=(0.5, 0.5),
                    phonemes_noise=0.5) :
        self.most_probable_phonemes = most_probable_phonemes
        self.phonemes_order_noise = phonemes_order_noise
        self.speakers_noise = speakers_noise
        self.phonemes_noise = phonemes_noise
    
    def _order_noise(self, sequence: List[str]) -> str :
        """
        Making noise the order of the phonemes in a given sequence

        Parameters
        ----------
        - sequence : list
            The sequence for which the phonemes order must be noised.
        
        Returns
        -------
        - str :
            The sequence with the order of phonemes noised.
        """
        # number of phonemes to noise in the sequence = len(sequence) / nb_phonemes_to_noise
        phonemes_to_noise = round(len(sequence) * self.phonemes_order_noise)
        # sample nb_phonemes_to_noise positions in the sequence
        positions_sampled = list(sample(range(len(sequence)), k=phonemes_to_noise))
        copied_positions = copy.deepcopy(positions_sampled)
        shuffle(copied_positions)
        # change the positions of the sampled phonemes 
        for original_position, new_position in zip(positions_sampled, copied_positions):
            sequence[original_position] = sequence[new_position]
        return " ".join(sequence)

    def _phonemes_noise(self, sequence: List[str]) -> str :
        """
        Makinng noise the phonemes of the sequence by replacing\
        some phonemes of the sequence by arbitrary sampled phonemes\
        from the vocabulary.

        Parameters
        ----------
        - sequence : list
            The sequence for which the phonemes must be noised.
        
        Returns
        -------
        - str :
            The sequence with noised phonemes.
        """
        phonemes_to_noise = round(len(sequence) * self.phonemes_noise)
        indexes = choices(range(len(sequence)), k=phonemes_to_noise)
        # choose new phonemes only from the most probable phonemes.
        phonemes = choices(self.most_probable_phonemes, k=phonemes_to_noise)
        # and replace some indices of the sequence by those choosen phonemes
        for idx, phonemes in zip(indexes, phonemes) :
            sequence[idx] = phonemes
        return " ".join(sequence)
    
    def _speakers_noise(self, speakers_sequences: Dict[str, set]) -> Dict[str, set] :
        """
        Making noise in the speaker's statements.

        Parameters
        ----------
        - speakers_sequences : dict
            Dictionary containing the utterances for each speaker.
        
        Returns
        -------
        - dict :
            The dictionary containing the few statements interchanged between the two speakers.
        """
        first_speaker, second_speaker = "Target_Child", "Adult"
        noise_first_speaker, noise_second_speaker = self.speakers_noise
        speakers_sequences[second_speaker] = set(speakers_sequences[second_speaker])
        speakers_sequences[first_speaker] = set(speakers_sequences[first_speaker])
        # sample some percentage of utterances from each speaker
        sequences_to_noise_second_speaker = round(len(speakers_sequences[second_speaker]) * noise_second_speaker)
        sequences_to_noise_first_speaker = round(len(speakers_sequences[first_speaker]) * noise_first_speaker)
        sequences_noise_second_speaker = sample(list(speakers_sequences[second_speaker]), k=sequences_to_noise_second_speaker)
        sequences_noise_first_speaker = sample(list(speakers_sequences[first_speaker]), k=sequences_to_noise_first_speaker)
        # noise by interchanging sampled utterances
        speakers_sequences[second_speaker] -= set(sequences_noise_second_speaker)
        speakers_sequences[first_speaker] -= set(sequences_noise_first_speaker)
        speakers_sequences[second_speaker] |= set(sequences_noise_first_speaker)
        speakers_sequences[first_speaker] |= set(sequences_noise_second_speaker)
        # set to list
        speakers_sequences[first_speaker] = list(speakers_sequences[first_speaker])
        speakers_sequences[second_speaker] = list(speakers_sequences[second_speaker])
        
        return speakers_sequences
    
    def __call__(self, loaded_dataset: dict) -> dict:
        """
        Apply the three types of noise.

        Parameters
        ----------
        loaded_dataset : dict
            The dictionary containing the utterances for each family, at each and for each speaker.
        
        Returns
        -------
        dict :
            The noised data.
        """
        for family in loaded_dataset :
            for age in loaded_dataset[family] :
                if "Adult" in loaded_dataset[family][age] :
                    for idx, utterance in enumerate(loaded_dataset[family][age]["Adult"]) :
                        splitted_utterance = utterance.split(" ")
                        loaded_dataset[family][age]["Adult"][idx] = self._order_noise(splitted_utterance)
                        loaded_dataset[family][age]["Adult"][idx] = self._phonemes_noise(splitted_utterance)
                if "Target_Child" in loaded_dataset[family][age] :
                    for idx, utterance in enumerate(loaded_dataset[family][age]["Target_Child"]) :
                        splitted_utterance = utterance.split(" ")
                        loaded_dataset[family][age]["Target_Child"][idx] = self._order_noise(splitted_utterance)
                        loaded_dataset[family][age]["Target_Child"][idx] = self._phonemes_noise(splitted_utterance)
                if "Target_Child" in loaded_dataset[family][age] and "Adult" in loaded_dataset[family][age] and all(self.speakers_noise):
                    noised_speaker = self._speakers_noise(loaded_dataset[family][age])
                    loaded_dataset[family][age]["Target_Child"] = noised_speaker["Target_Child"]
                    loaded_dataset[family][age]["Adult"] = noised_speaker["Adult"]
        return loaded_dataset