Method and apparatus for training language model and recognizing speech

Patent metadata

Field	Value
Publication number	US-9666184-B2
Application number	US-201514727462-A
Country	US
Kind code	B2
Filing date	Jun 1, 2015
Priority date	Dec 8, 2014
Publication date	May 30, 2017
Grant date	May 30, 2017

Abstract

A method and apparatus for training a neural network language model, and a method and apparatus for recognizing speech data based on a trained language model are provided. The method of training a language model involves converting, using a processor, training data into error-containing training data, and training a neural network language model using the error-containing training data.

First claim

What is claimed is:

1. A method of training a language model, the method comprising:
converting, using a processor, training data into error-containing training data; and
training a neural network language model using the error-containing training data
wherein the converting comprises selecting a word to be replaced with an erroneous word from words in the training data, and generating the error-containing training data by replacing the selected word with the erroneous word,
wherein the neural network language model is used to estimate a connection relationship between words,
wherein the selecting comprises randomly selecting the word from the words in the training data,
wherein the processor is configured to use the trained language model to convert a speech into output data.

2. The method of claim 1 , wherein the converting further comprises selecting the erroneous word from a plurality of candidate words associated with the selected word.

3. The method of claim 2 , wherein the candidate words are determined based on phonetic similarities to the selected word.

4. The method of claim 2 , wherein the selecting of the erroneous word comprises selecting the erroneous word based on weights assigned to the candidate words.

5. The method of claim 1 , further comprising:
determining whether the training data is to be converted into the error-containing training data.

6. The method of claim 1 , wherein the neural network language model generates a probability value.

7. A non-transitory computer-readable storage medium comprising a program comprising instructions to cause a computer to perform the method of claim 1 .

8. An apparatus for training a language model, the apparatus comprising:
a processor configured to convert training data into error-containing training data, and train the neutral network language model using the error-containing training data,
wherein the processor is configured to select a word to be replaced with an erroneous word from words in the training data, and generate the error-containing training data by replacing the selected word with the erroneous word,
wherein the neural network language model is used to estimate a connection relationship between words,
wherein the processor is configured to randomly select the word from the words in the training data,
wherein the processor is configured to use the trained language model to convert a speech into output data.

9. The apparatus of claim 8 , wherein the processor is configured to select the erroneous word from a plurality of candidate words determined based on phonetic similarities to the selected word.

10. The apparatus of claim 8 , wherein the processor is configured to determine whether the training data is to be converted into the error-containing training data.

11. The apparatus of claim 8 , wherein the processor generates the error-containing training data by retrieving the training data from a memory storage and replacing a selected word included in the training data with an erroneous word selected based on phonetic similarity of the erroneous word to the selected word.

12. The apparatus of claim 8 , wherein the processor is configured to use the trained language model to convert a speech received from a microphone into output data.

13. The apparatus of claim 12 , wherein the apparatus is configured to use the output data as a user command in controlling applications on the apparatus.

14. A method of training a language model, the method comprising:
converting, using a processor, training data into error-containing training data; and
training a neural network language model using the error-containing training data,
wherein the converting comprises selecting a word to be replaced with an erroneous word from words in the training data and generating the error-containing training data by replacing the selected word with the erroneous word,
wherein the neural network language model is trained by acoustically embedding the error-containing training data in a projection layer of the neural network,
wherein the processor is configured to use the trained language model to convert a speech into output data.

Assignees

• Samsung Electronics Co Ltd

Inventors

• Lee Hoshik
• Choi Heeyoul

Classifications

• G10L15/16 Primary

  using artificial neural networks · CPC title

• G10L15/187

  Phonemic context, e.g. pronunciation rules, phonotactical constraints or phoneme n-grams · CPC title

• G10L15/19

  Grammatical context, e.g. disambiguation of the recognition hypotheses based on word sequence rules · CPC title

• G10L25/30

  using neural networks · CPC title

• G10L15/063

  Training · CPC title