By using rules method we can define custom behaviour and add necessary rules within it. Like symbols and abbreviations.
We have to provide the rules data and update the vocabulary accordingly.
The rule_data is a dictionary like {"dollar" : '$', "C M" : "CM"} etc.
We can also include regex or functions as dictionary elements. Like lambda x:x*3 for triple.
| class Speech2text(tf.keras.Model): | |
| def __init__(self, vocab_size, lstm_units, rules): | |
| super().__init__() | |
| self.vocab_size = vocab_size | |
| self.lstm_units= lstm_units | |
| self.rule_data = rules # a dictionary | |
| self.dense = Dense(vocab_size, activation="softmax") | |
| def build(self, input_shape): | |
| self.lstm = LSTM(self.lstm_units, return_state=True, return_sequences=True) | |
| def rules(word): | |
| # can add any extra functionality through this method; like symbols, abbreviation etc. | |
| if word in self.rule_data: | |
| return self.rule_data[word] | |
| else: | |
| return word | |
| def call(self, audio, states): | |
| lstm_output, lstm_state_h, lstm_state_c = self.lstm(audio) | |
| id = dense(self.lstm_output) | |
| word = tokenizer.id_to_txt(id) | |
| word = self.rules(word) | |
| return lstm_output, lstm_state_h, lstm_state_c, word |