Skip to content

g2p

GraphemeToPhonemeTokenizer

Bases: Tokenizer

Tokenizer implementation that converts the input sentence string into phoneme tokens by the g2p package.

References: https://github.com/Kyubyong/g2p

Source code in speechain/tokenizer/g2p.py 
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
class GraphemeToPhonemeTokenizer(Tokenizer):
    """Tokenizer implementation that converts the input sentence string into phoneme
    tokens by the g2p package.

    References: https://github.com/Kyubyong/g2p
    """

    def text2tensor(
        self,
        text: str or List[str],
        no_sos: bool = False,
        no_eos: bool = False,
        return_tensor: bool = True,
    ):
        """
        This text-to-tensor function can take two types of input:
        1. raw string of the transcript sentence
        2. structured string of the phonemes dumped in advance

        But we recommend you to feed the type no.2 to this function because if the input it type no.1, the raw string
        needs to be decoded by g2p_en.G2p in each epoch, which not only consumes a lot of CPU but also slow down the
        model forward.

        """
        # initialize the tensor as an empty list
        tokens = []
        # whether to attach sos at the beginning of the tokens
        if not no_sos:
            tokens.append(self.sos_eos_idx)

        # when input text is a dumped phoneme list
        if isinstance(text, List):
            tokens += [
                (
                    self.token2idx[token]
                    if token in self.token2idx.keys()
                    else self.unk_idx
                )
                for token in text
            ]
        # when input text is a raw string
        else:
            # initialize g2p convertor lazily during training
            if not hasattr(self, "g2p"):
                self.g2p = G2p()
            phonemes = self.g2p(text)
            for phn in phonemes:
                if phn in abnormal_phns:
                    continue
                elif phn == " ":
                    tokens.append(self.space_idx)
                elif phn not in self.token2idx.keys():
                    tokens.append(self.unk_idx)
                else:
                    tokens.append(self.token2idx[phn])

        # whether to attach eos at the end of the tokens
        if not no_eos:
            tokens.append(self.sos_eos_idx)

        if return_tensor:
            return torch.LongTensor(tokens)
        else:
            return tokens

text2tensor(text, no_sos=False, no_eos=False, return_tensor=True)

This text-to-tensor function can take two types of input: 1. raw string of the transcript sentence 2. structured string of the phonemes dumped in advance

But we recommend you to feed the type no.2 to this function because if the input it type no.1, the raw string needs to be decoded by g2p_en.G2p in each epoch, which not only consumes a lot of CPU but also slow down the model forward.

Source code in speechain/tokenizer/g2p.py 
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
def text2tensor(
    self,
    text: str or List[str],
    no_sos: bool = False,
    no_eos: bool = False,
    return_tensor: bool = True,
):
    """
    This text-to-tensor function can take two types of input:
    1. raw string of the transcript sentence
    2. structured string of the phonemes dumped in advance

    But we recommend you to feed the type no.2 to this function because if the input it type no.1, the raw string
    needs to be decoded by g2p_en.G2p in each epoch, which not only consumes a lot of CPU but also slow down the
    model forward.

    """
    # initialize the tensor as an empty list
    tokens = []
    # whether to attach sos at the beginning of the tokens
    if not no_sos:
        tokens.append(self.sos_eos_idx)

    # when input text is a dumped phoneme list
    if isinstance(text, List):
        tokens += [
            (
                self.token2idx[token]
                if token in self.token2idx.keys()
                else self.unk_idx
            )
            for token in text
        ]
    # when input text is a raw string
    else:
        # initialize g2p convertor lazily during training
        if not hasattr(self, "g2p"):
            self.g2p = G2p()
        phonemes = self.g2p(text)
        for phn in phonemes:
            if phn in abnormal_phns:
                continue
            elif phn == " ":
                tokens.append(self.space_idx)
            elif phn not in self.token2idx.keys():
                tokens.append(self.unk_idx)
            else:
                tokens.append(self.token2idx[phn])

    # whether to attach eos at the end of the tokens
    if not no_eos:
        tokens.append(self.sos_eos_idx)

    if return_tensor:
        return torch.LongTensor(tokens)
    else:
        return tokens