Language model customization in speech recognition for speech analytics

Patent metadata

Field	Value
Publication number	US-10643604-B2
Application number	US-201816219537-A
Country	US
Kind code	B2
Filing date	Dec 13, 2018
Priority date	Jan 16, 2016
Publication date	May 5, 2020
Grant date	May 5, 2020

Abstract

A method for generating a language model for an organization includes: receiving, by a processor, organization-specific training data; receiving, by the processor, generic training data; computing, by the processor, a plurality of similarities between the generic training data and the organization-specific training data; assigning, by the processor, a plurality of weights to the generic training data in accordance with the computed similarities; combining, by the processor, the generic training data with the organization-specific training data in accordance with the weights to generate customized training data; training, by the processor, a customized language model using the customized training data; and outputting, by the processor, the customized language model, the customized language model being configured to compute the likelihood of phrases in a medium.

First claim

What is claimed is:

1. A method for performing voice analytics on interactions with an organization, comprising:
training a customized language model for the organization by:
receiving, by a speech recognition engine, organization-specific training data and generic training data;
computing, by the speech recognition engine, a plurality of similarities between the generic training data and the organization-specific training data;
assigning, by the speech recognition engine, a plurality of weights to the generic training data through partitioning the generic training data into a plurality of partitions in accordance with the computed similarities wherein the computed similarities comprise a fixed set of one or more threshold similarities, associating a partition similarity with each of the partitions, the partition similarity corresponding to the average similarity of the data in the partition, and assigning a desired weight to each partition, the desired weight corresponding to the partition similarity of the partition;
combining, by the speech recognition engine, the generic training data with the organization-specific training data in accordance with the weights to generate customized training data;
training, by the speech recognition engine, the customized language model using the customized training data; and
outputting, by the speech recognition engine, the customized language model, the customized language model being configured to compute a likelihood of phrases in a medium;

receiving, by the speech recognition engine, an input speech from an interaction between a customer and an agent of the organization; and
performing voice analytics on the received input speech.

2. The method of claim 1 , wherein a silhouette score is used to determine a number of the plurality of partitions.

3. The method of claim 1 , wherein a test set of the generic training data and the organization-specific training data empirically determine a number of the plurality of partitions.

4. The method of claim 1 , wherein k-means clustering is used to determine a number of the plurality of partitions.

5. The method of claim 1 , wherein the desired weight of a partition is exponentially decreasing with decreasing partition similarity.

6. The method of claim 1 , wherein the training a customized language model for the organization further comprise:
receiving organization-specific in-medium data;
combining the organization-specific in-medium data with the generic training data and the organization-specific training data to generate the customized training data; and
retraining the language model in accordance with the customized training data.

7. The method of claim 1 , wherein the organization-specific training data comprise at least one of: in-medium data and out-of-medium data.

8. The method of claim 7 , wherein the in-medium data comprise speech recognition transcript text and the out-of-medium data comprise non-speech text.

9. A voice analytics system comprising:
a speech model training system comprising:
a processor; and
memory coupled to the processor and storing instructions that, when executed by the processor, cause the processor to:
receive organization-specific training data and generic training data;
compute a plurality of similarities between the generic training data and the organization-specific training data;
assign a plurality of weights to the generic training data through partitioning the generic training data into a plurality of partitions in accordance with the computed similarities wherein the computed similarities comprise a fixed set of one or more threshold similarities, associating a partition similarity with each of the partitions, the partition similarity corresponding to the average similarity of the data in the partition, and assigning a desired weight to each partition, the desired weight corresponding to the partition similarity of the partition;
combine the generic training data with the organization-specific training data in accordance with the weights to generate customized training data;
train a customized language model using the customized training data; and
output the customized language model, the customized language model being configured to compute the likelihood of phrases in a medium; and

a speech analytics system configured to:
receive an input speech from an interaction between a customer and an agent of the organization; and
perform voice analytics on the received input speech.

10. The speech recognition system of claim 9 , wherein a silhouette score is used to determine a number of the plurality of partitions.

11. The speech recognition system of claim 9 , wherein a test set of the generic training data and the organization-specific training data empirically determine a number of the plurality of partitions.

12. The speech recognition system of claim 9 , wherein k-means clustering is used to determine a number of the plurality of partitions.

13. The speech recognition system of claim 9 , wherein the desired weight of a partition is exponentially decreasing with decreasing partition similarity.

14. The speech recognition system of claim 9 , wherein the memory of the speech training model system further stores instructions that, when executed by the processor, cause the processor to:
receive organization-specific in-medium data;
combine the organization-specific in-medium data with the generic training data and the organization-specific training data to generate the customized training data; and
retrain the language model in accordance with the customized training data.

15. The speech recognition system of claim 9 , wherein the organization-specific training data comprise at least one of: in-medium data and out-of-medium data.

16. The speech recognition system of claim 15 , wherein the in-medium data comprise speech recognition transcript text and the out-of-medium data comprise non-speech text.

Assignees

• Genesys Telecommunications Laboratories Inc

Inventors

• Tapuhi Tamir
• Lev-Tov Amir
• Faizakof Avraham
• Konig Yochai

Classifications

• G10L15/063 Primary

  Training · CPC title

• G10L15/183

  using context dependencies, e.g. language models · CPC title

• G10L2015/088

  Word spotting · CPC title

• G10L2015/0636

  Threshold criteria for the updating · CPC title

• G06N3/006

  based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO] · CPC title