JPS63163495A - Voice section detector - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION Page 2 Industrial Application Field The present invention relates to a speech segment detection device used in speech recognition devices and the like.

BACKGROUND OF THE INVENTION In recent years, with the advancement of voice recognition technology, various devices using voice recognition have begun to be created. However, one of the problems in recognition technology is accurately extracting speech sections from the input signal, and various speech section detection devices have been developed (for example, Japanese Patent Application Laid-open No. 6゜-52900). .

A conventional voice section detection device will be explained below. Second
The figure is a block diagram of a conventional voice section detection device. 1 is an audio input terminal, 8 is a BPF (band pass filter) that extracts the characteristics of the audio signal and cuts noise, and 5HBPF.
an energy level comparator that determines a voice section by comparing the energy level passed through s with a preset threshold;
2 is a voice section output terminal, 9 (l-I voice section output terminal 2
This is a voice recognition matching unit that outputs recognition results from the voice signals output from the voice signals.

The operation of the conventional voice section detection device configured as 3 or more will be described below.

First, an audio signal is input from the audio input terminal 1 and passes through BPFs whose bands are set in advance.

The band of BPF8 is set to a band that clearly shows the characteristics of the voice and does not cause noise. The audio signal that has passed through the BPF 8 is compared in the energy level comparator 3 with the energy of the audio signal and a preset threshold, and the section where the energy is greater than the threshold is regarded as a speech section and is sent to the speech section output terminal 2. Output. The speech recognition matching section compares the speech signal outputted from the speech section output terminal 2 with a standard pattern of speech stored in advance and outputs a recognition result.

Problems to be Solved by the Invention However, in the conventional configuration described above, speech sections are detected using energy in a frequency band excluding the noise frequency band. ) was being mistakenly detected as voice. It was still difficult to set a band with voice characteristics.

SUMMARY OF THE INVENTION The present invention solves the above-mentioned conventional problems, and aims to provide a speech section detection device that can accurately separate speech sections from noise sections.

Means for Solving the Problem In order to achieve this objective, the speech section detection device of the present invention includes an energy level comparator that compares the energy of the speech input with a preset threshold and detects a temporary speech section. and a noise detector that detects whether noise is included in the temporary voice section detected by the energy level comparator and removes the noise section if noise is included.

Note that the noise detector includes a pitch randomness calculation means for calculating the temporal randomness of the pitch frequency of the audio input, and a preset threshold value and the pinch randomness calculated by the pitch randomness calculation means. In comparison, when the duration of the state in which the randomness of the pinch is greater than the threshold is longer than the preset threshold time, and the audio input that has passed through the low-pass filter 5 is equal to or greater than the preset threshold, this section is considered to be noise. It is comprised of a noise determining means for making a determination, and a noise section removing means for removing the noise section detected by the noise determining means from a temporary speech section.

Operation According to the above-described configuration, the present invention first detects a temporary speech section by comparing the energy of the speech input with a preset threshold using the energy level comparator. Next, the pitch randomness calculation means of the noise detector calculates the randomness of the pinch frequency of the audio input in the time direction. When the randomness of this pinch is greater than the preset threshold in the noise judgment means of the noise detector and continues for more than the preset threshold time, and the energy below the preset frequency is greater than the preset threshold, this The section is determined to be noise. Then, the noise section removal means of the noise detector removes the noise section from the temporary speech section and outputs the result, so that the speech section can be extracted accurately.

6 BENO EMBODIMENT An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram of a voice section detection device according to an embodiment of the present invention. In FIG. 1, 1 is an audio input terminal, 3 is an energy level comparator that compares the energy level of the audio input from the audio input terminal 1 with a preset threshold value, and detects a temporary audio section; an LPF that passes only signals of 400 or less of the audio input; 5 a pitch detector that detects the pitch frequency of the audio input; 71 a pitch randomness that calculates the randomness of the pitch frequency detected by the pitch detector 5; A calculator 72 determines the pitch randomness calculated by the pitch randomness calculation means 71 and the energy of the voice input through the LPF 4 for the temporary voice section detected by the energy level comparator 3. noise determining means for determining whether the noise is noise or voice; 7
Reference numeral 3 denotes a noise section removing means for removing the section determined to be noise by the noise determining means 72. 2 to voice section 7
The inter-Beno output terminal 9 is a voice recognition matching section, and these are the same as in the conventional example. The noise detector 7 is composed of a pitch randomness calculating means 71, a noise determining means 72, and a noise section removing means 73.

The operation of the voice section detection device of this embodiment configured as described above will be explained below.

First, when a voice input is input from the voice input terminal 1, the energy level comparator 3 compares the energy of the voice input with a preset threshold, and determines a section where the energy is greater than the threshold as a temporary voice section. I reckon. Also, only signals of 4001-Iz or less are passed through the audio input HLPF 4, while a pinch frequency is calculated by the pitch detector 5, and randomness is added to the pitch frequency by a pitch randomness calculating means 71. Calculated.

Next, one embodiment of the pitch randomness calculation means 71 will be explained. One frame (
For example, let it be the randomness of the pitch of the frame with a 5-point median e4 of the difference between the frames before and after the pitch frequency value every 12 m58Q). Next, the noise determining means 72 determines that the state in which the pitch randomness is equal to or greater than a predetermined threshold continues for a predetermined number of frames or more, and the energy of the audio input passed through the LPF 4 is equal to or greater than the predetermined threshold. In this case, the section with large pitch randomness is regarded as a noise section, and the noise section removing means 73
Then, the noise section is removed from the temporary speech section detected by the energy level comparator 3, and outputted to the speech section output terminal 2, which is input to the speech recognition matching section 9.

As described above, according to this embodiment, the noise detector 7 detects the noise included in the temporary speech section detected by the energy level comparator 3, and if noise is present, the noise section is removed. , accurate speech intervals can be detected. Furthermore, the noise detector 7 is equipped with a noise determining means 72 that detects noise based on the randomness of the pitch, its duration, and the energy of the audio signal passing through the LPF, so that it can detect noise with high accuracy. Since the LPF 4 and the pitch detector 6 are still used for extracting feature patterns for speech recognition, there is no need to add them for the speech section detection device of the present invention, and there is no need to add them in terms of device size, cost, and processing speed. This is an excellent voice section detection device.

Effects of the Invention By providing a noise detector, the present invention can detect noise included in the voice section detected by the energy comparator and remove the noise section when noise is present. detects noise based on the randomness of the pinch, its duration, and the energy of the audio signal passing through the LPFi, so it can detect noise with high accuracy.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of a speech segment detection device according to an embodiment of the present invention, and FIG. 2 is a block diagram of a conventional speech segment detection device. 1...Audio input terminal, 2...Audio section output terminal, 3...Energy level comparator, 4.
...LPF, 5...Pitch detector, 7.
... Noise detector, 71 ... Pitch randomness calculation means, 72 ... Noise judgment 1 o vane means, 73 ... Noise section removal means, 9・・・・・・
...Speech recognition matching section, 8...BPF.

Claims (1)

[Claims] an energy level comparator that detects a temporary voice section by comparing the energy of the voice input with a preset threshold; a pitch randomness calculation means that calculates the randomness of the pitch frequency of the voice input in the time direction; The pitch randomness calculated by the pitch randomness calculation means is compared with a preset threshold value, and the duration of the state in which the pitch randomness is greater than the threshold value is equal to or longer than the preset threshold time, and the low-pass filter is Three means: a noise determining means that determines this section as noise when the voice input passed through it is equal to or higher than a preset threshold, and a noise section removing means that removes the noise section detected by the noise determining means from the temporary voice section. What is claimed is: 1. A speech interval detection device comprising: a noise detector comprising: