Device, method, and computer program product for executing inference using input signal

Patent metadata

Field	Value
Publication number	US-12437215-B2
Application number	US-202016942906-A
Country	US
Kind code	B2
Filing date	Jul 30, 2020
Priority date	Jan 28, 2020
Publication date	Oct 7, 2025
Grant date	Oct 7, 2025

Abstract

This signal processing device includes one or more processors. The processors receive, as an input, an input signal that is a third signal obtained by superposing a second signal on a first signal or a fourth signal obtained by converting the third signal, and estimate a feature of the first signal on the basis of the input signal. The processors execute inference on the basis of the feature and outputs an inference result.

First claim

What is claimed is:

1. A signal processing device, comprising:
one or more processors configured to:
acquire, from a plurality of outputs, each of which is an output of a plurality of first learning models for learning, a feature of a first signal, that is an audio signal of an utterance of a speaker used for inference, by applying an input signal to each of the first learning models for learning such that the acquired feature is outputted upon inputting the input signal, the input signal being a third signal or a fourth signal, the third signal including the first signal and a second signal that is a signal unnecessary for the inference, the fourth signal being obtained by converting the third signal, wherein the acquired feature is frequency information representing respective frequencies of a plurality of signals contained in the first signal;
display, on a display, the acquired feature by using the plurality of first learning models and the identity of the word spoken by the speaker, wherein the display comprises an interactive slide bar for designating one or more weighting factors that are used to add together, based on the one or more weighting factors, a plurality of respective features obtained based on each of the plurality of first learning models and output the addition result as the acquired feature of the first signal, and wherein designating the one or more weighting factors updates the acquired feature displayed on the display;
execute inference by using a second learning model for learning such that an inference result is outputted upon inputting the acquired feature, the inference result being an indication of an identity of a word spoken by the speaker;
calculate both a first error value and a second error value, the first error value constituting an error value between a first correct answer signal representing a correct answer of the acquired feature and the acquired feature, the second error value constituting an error value between a second correct answer signal representing a correct answer of inference based on the acquired feature and the outputted inference result; and
execute a training process by executing both (1) first learning processing to update a parameter of each of the plurality of first learning models based on both the first error value and the second error value, and (2) second learning processing to update a parameter of the second learning model based on the second error value, wherein the first learning processing to update the parameter of each of the plurality of first learning models includes:
performing a multiplying process that multiplies the first error value by a first adjustment factor and multiplies the second error value by a second adjustment factor;
updating the parameter of the one or more first learning models based on a sum of the first error value and the second error value after the multiplying process; and
modifying a value of the first adjustment factor and a value of the second adjustment factor such that, as a number of times of updating the parameter of the one or more first learning models increases, the first error value after being multiplied by the first adjustment factor is reduced and the second error value after being multiplied by the second adjustment factor is increased with the number of times of updating.

2. The signal processing device according to claim 1 , wherein the one or more processors are further configured to set the first adjustment factor and the second adjustment factor such that the first error value after the multiplying process is greater than the second error value.

3. The signal processing device according to claim 1 , wherein the parameter of each of the plurality of first learning models is having been updated based on each of a plurality of first error values based on a plurality of different indices.

4. The signal processing device according to claim 3 , wherein the one or more processors are further configured to acquire the feature by using a learning model obtained by adding together the plurality of first learning models based on the one or more weighting factors.

5. The signal processing device according to claim 1 , wherein the one or more processors are further configured to estimate the acquired feature by using a learning model obtained by adding together the plurality of first learning models based on the one or more weighting factors.

6. The signal processing device according to claim 1 , further comprising:
a memory configured to store a plurality of stored data associating the acquired feature with the inference result, wherein
the one or more processors are further configured to read, from the memory, a predetermined number of the stored data associating features that are selected in descending order of error value with respect to the acquired feature, and display, on the display, the stored data thus read.

7. The signal processing device according to claim 6 , wherein the one or more processors are further configured to match a coordinate axis displaying a particular feature contained in the stored data with a coordinate axis displaying the acquired feature.

8. The signal processing device according to claim 1 , wherein:
the one or more processors are further configured to determine, based on the acquired feature, whether or not inference is to be executed, and
execute the inference in a case where a determination to execute the inference has been made.

9. A signal processing method, comprising:
acquiring, from a plurality of outputs, each of which is an output of a plurality of first learning models for learning, a feature of a first signal, that is an audio signal of an utterance of a speaker used for inference, by applying an input signal to each of the one or more first learning models for learning such that the acquired feature is outputted upon inputting the input signal, the input signal being a third signal or a fourth signal, the third signal including the first signal and a second signal that is a signal unnecessary for the inference, the fourth signal being obtained by converting the third signal, wherein the acquired feature is frequency information representing respective frequencies of a plurality of signals contained in the first signal;
displaying, on a display, the acquired feature by using the plurality of first learning models and the identity of the word spoken by the speaker, wherein the display comprises an interactive slide bar for designating one or more weighting factors that are used to add together, based on the one or more weighting factors, a plurality of respective features obtained based on each of the plurality of first learning models and output the addition result as the acquired feature of the first signal, and wherein designating the one or more weighting factors updates the acquired feature displayed on the display;
executing inference by using a second learning model for learning such that an inference result is outputted upon inputting the acquired feature, the inference result being an indication of an identity of a word spoken by the speaker;
calculating both a first error value and a second error value, the first error value constituting an error value between a first correct answer signal representing a correct answer of the acquired feature and the acquired feature, the second error value constituting an error value between a second correct answer signal representing a correct answer of inference based on the acquired feature and the outputted inference result; and
executing a training process by executing both (1) first learning processing to update a parameter of each of the plurality of first learning models based on both the first error value and the second error value, a

Assignees

• Toshiba Kk
• Toshiba Infrastructure Systems & Solutions Corp

Inventors

• Sasaya Tenta
• Watanabe Takashi
• Ono Toshiyuki

Classifications

• G06N20/00

  Machine learning · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N5/04 Primary

  Inference or reasoning models · CPC title

• G06N3/08 Primary

  Learning methods · CPC title