Computer-implemented method and apparatus for generating grapheme-to-phoneme model

• US 10,181,320 B2
• Filed: 12/28/2016
• Issued: 01/15/2019
• Est. Priority Date: 02/24/2016
• Status: Active Grant

• Current Assignee: Baidu Online Network Technology (Beijing) Co., Ltd (Baidu Incorporated)
• Original Assignee: Baidu Online Network Technology (Beijing) Co., Ltd (Baidu Incorporated)
• Inventors: Chen, Zhijie, Kang, Yongguo
• Primary Examiner(s): Pham, Ly D
• Application Number:
  US15/391,907
• Publication Number:
  US 20170243575A1
• Time in Patent Office: 748 Days
• Field of Search: 706 25
• CPC Class Codes: G06F 40/126 Character encoding
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G06N 3/044 Recurrent networks, e.g. Hopfield networks
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G10L 13/00 Speech synthesis; Text to speech systems
• CPC Class Codes: G10L 15/02 Feature extraction for speech recognition; Selection of recognition unit
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/16 using artificial neural networks
• CPC Class Codes: G10L 2015/025 Phonemes, fenemes or fenones being the recognition units

Abstract

A method and an apparatus for generating a g2p model based on AI are provided. The method includes: during performing a grapheme-to-phoneme conversion training by a neural network on each word in training data, screening nodes in a hidden layer of the neural network randomly according to a preset node ratio so as to obtain retaining nodes for training each word; training each word with a sub-neural network corresponding to the retaining nodes and updating a weight of each retaining node of the sub-neural network; and performing a mean processing on the weights of the retaining nodes of respective sub-neural networks, so as to generate the grapheme-to-phoneme model.

Claims

1. A computer-implemented method for generating a grapheme-to-phoneme model based on artificial intelligence, performed by one or more computing devices and comprising:

during performing a grapheme-to-phoneme conversion training by a neural network on each word in training data, screening nodes in a hidden layer of the neural network randomly according to a preset node ratio so as to obtain retaining nodes for training each word;

training each word with a sub-neural network corresponding to the retaining nodes and updating a weight of each retaining node of the sub-neural network;

performing a mean processing on the weights of the retaining nodes of respective sub-neural networks, so as to generate the grapheme-to-phoneme model; and

adjusting the number of layers in the neural network, the number of nodes in each layer and the preset node ratio according to the number of words in the training data.

2. The method according to claim 1, wherein screening nodes in a hidden layer of the neural network randomly according to a preset node ratio so as to obtain retaining nodes for training each word comprises:

hiding a portion of nodes in each hidden layer of the neural network randomly according to a preset hiding ratio of nodes, such that a portion of nodes unhidden in each hidden layer of the neural network are used as the retaining nodes for training each word, in which the number of the nodes hidden corresponds to the preset hiding ratio of nodes.

3. The method according to claim 1, wherein screening nodes in a hidden layer of the neural network randomly according to a preset node ratio so as to obtain retaining nodes for training each word comprises:

retaining a portion of nodes in each hidden layer of the neural network randomly according to a preset retaining ratio of nodes, and the nodes retained are used as the retaining nodes for training each word, in which the number of the nodes retained corresponds to the preset retaining ratio of nodes.

4. A computer-implemented apparatus for generating a grapheme-to-phoneme model based on artificial intelligence, comprising one or more computing devices configured to execute one or more software modules, the one or more software modules comprising:

a processing module, configured to screen nodes in a hidden layer of a neural network randomly according to a preset node ratio so as to obtain retaining nodes for training each word, during performing a grapheme-to-phoneme conversion training by the neural network on each word in training data;

a training module, configured to train each word with a sub-neural network corresponding to the retaining nodes and to update a weight of each retaining node of the sub-neural network; and

a generating module, configured to perform a mean processing on the weights of the retaining nodes of respective sub-neural networks, so as to generate the grapheme-to-phoneme model;

wherein the processing module is further configured to;

adjust the number of layers in the neural network, the number of nodes in each layer and the preset node ratio according to the number of words in the training data.

5. The apparatus according to claim 4, wherein the processing module is configured to:

hide a portion of nodes in each hidden layer of the neural network randomly according to a preset hiding ratio of nodes, such that a portion of nodes unhidden in each hidden layer of the neural network are used as the retaining nodes for training each word, in which the number of the nodes hidden corresponds to the preset hiding ratio of nodes.

6. The apparatus according to claim 4, wherein the processing module is configured to:

retain a portion of nodes in each hidden layer of the neural network randomly according to a preset retaining ratio of nodes, and the nodes retained are used as the retaining nodes for training each word, in which the number of the nodes retained corresponds to the preset retaining ratio of nodes.

7. A non-transitory computer-readable storage medium having stored therein instructions that, when executed by a processor of a device, causes the device to perform a computer-implemented method for generating a grapheme-to-phoneme model based on artificial intelligence, the computer-implemented method comprising:

during performing a grapheme-to-phoneme conversion training by a neural network on each word in training data, screening nodes in a hidden layer of the neural network randomly according to a preset node ratio so as to obtain retaining nodes for training each word;

training each word with a sub-neural network corresponding to the retaining nodes and updating a weight of each retaining node of the sub-neural network; and

performing a mean processing on the weights of the retaining nodes of respective sub-neural networks, so as to generate the grapheme-to-phoneme model;

adjusting the number of layers in the neural network, the number of nodes in each layer and the preset node ratio according to the number of words in the training data.

8. The non-transitory computer-readable storage medium according to claim 7, wherein screening nodes in a hidden layer of the neural network randomly according to a preset node ratio so as to obtain retaining nodes for training each word comprises:

hiding a portion of nodes in each hidden layer of the neural network randomly according to a preset hiding ratio of nodes, such that a portion of nodes unhidden in each hidden layer of the neural network are used as the retaining nodes for training each word, in which the number of the nodes hidden corresponds to the preset hiding ratio of nodes.

9. The non-transitory computer-readable storage medium according to claim 7, wherein screening nodes in a hidden layer of the neural network randomly according to a preset node ratio so as to obtain retaining nodes for training each word comprises:

retaining a portion of nodes in each hidden layer of the neural network randomly according to a preset retaining ratio of nodes, and the nodes retained are used as the retaining nodes for training each word, in which the number of the nodes retained corresponds to the preset retaining ratio of nodes.