JPH071433B2 - Voice start time detection method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION Object of the Invention (Industrial field of use) The present invention relates to a voice start time detection method for detecting a voice start time.

(Prior Art) A voice recognition device spectrally analyzes an input voice waveform and stores it as a voice input pattern in a pattern memory. This voice input pattern is compared with the standard pattern registered in the dictionary in advance to calculate the similarity, and the standard pattern with the highest similarity is determined and output as the recognition result.

Such voice recognition devices include a word voice recognition device that recognizes a voice in word units by an input voice and a monosyllabic voice recognition device that recognizes a voice in a monosyllabic unit. In both the word voice recognition device and the monosyllabic voice recognition device, it is necessary to determine the start time of the voice. Conventionally, it is simply detected whether or not the level of the voice waveform exceeds a predetermined value, and the time when the level exceeds a certain level is set as the voice start time.

(Problems to be Solved by the Invention) However, the conventional method has a problem in that noise and voice cannot be accurately discriminated. In particular, there is a problem in that it is difficult to distinguish sharp noises from consonants, the voice start time is mistakenly detected, and as a result, the voice is erroneously recognized.

The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a voice start time point detection method capable of accurately detecting the voice start time point.

[Structure of Invention]

(Means for Solving Problems) In order to achieve the above object, a voice start time point detection method according to the present invention uses a sampling value of an input voice waveform at a plurality of time points at fixed time intervals in a direction opposite to the time axis with reference to the current time point. A plurality of sampling values are extracted for each time, and the sum of sampling values, which is the sum of the absolute values of the plurality of sampled values, is calculated, and all of these 22 pulling value sums exceed a predetermined threshold value. In this case, the voice start time is determined based on the current time.

(Embodiment) A method for extracting a consonant of a monosyllabic voice using a method for detecting a voice start point according to an embodiment of the present invention will be described with reference to the flowchart of FIG. In this embodiment, sampling is 100 μsec
It is assumed that 128 sampling values Si are set as one frame and voice recognition is performed in one frame.

First, an interrupt is applied every 100 μsec, the voice waveform is sampled, and the sampling value Si is stored in the memory (step 10).

Next, calculation for determining the voice start time is performed (step 11). The sampling value at the time of the interrupt is set to S ₀ and 100
Assuming that the sampling value before × nμsec is S ₋ n, the sampling value sums R ₁ , R ₂ and R ₃ shown in the following equation are obtained.

For example, if m = 8 now, as shown in FIG. 2, the sampling value sum R ₁ is the sampling value S ₋₁₂₈ at the time point one frame before.
~ The sum of absolute values of S _-120 is shown, and the sum of sampling values R ₂ is 2
The sum of the absolute values of the sampling values S _{-256 to} S _-248 at the time point before the frame is shown, and the sum R ₃ of the sampling values is the sum of the absolute values of the sampling values S _{-384 to} S _-376 at the time point three frames before. ing.

Next, it is determined whether or not all of these sampling value sums R ₁ , R ₂ , R ₃ are larger than a predetermined threshold value R _TH (step 12). Threshold value even with one of sampling values R ₁ , R ₂ and R ₃
If it is smaller than R _TH , it is judged as noise and the process returns to step 10. The sum of all sampling values R ₁ , R ₂ , R ₃ is the threshold value R _TH
If it is larger, it is determined that it is an actual voice, and in step 13, the voice start time is determined. In the present embodiment, a time point three frames in the past based on the current time, that is, a sampling value S _-384
Is the voice start time.

Next, a predetermined number of frames, for example, 8 frames are cut out as a consonant portion with reference to this voice start time point (step 14).

As described above, according to the present embodiment, since it is determined whether or not the voice is voice in a relatively long period of 38.4 msec, it is possible to discriminate the voice and the sharp noise having a large amplitude value.

The present invention is not limited to the above embodiment, and various modifications can be made.
For example, any number of sampling values may be used to calculate the sum of sampling values. Further, the number of frames for determining whether or not it is the voice start time is not limited to three, and may be any number. Further, the number of frames past from the present time point is not limited to the time point of three frames past as the voice start time point.

Furthermore, although the above-described embodiment has been applied to the consonant extraction of monosyllabic voices, it goes without saying that the present invention can also be applied to detection of the start point of a word voice or continuous voice.

〔The invention's effect〕

As described above, according to the present invention, it is possible to accurately detect the voice start point.

[Brief description of drawings]

FIG. 1 is a flowchart of a method for detecting a voice start point according to an embodiment of the present invention, and FIG. 2 is a waveform diagram showing an input voice waveform.

Claims (1)

[Claims] 1. A plurality of sampling values of an input speech waveform are taken out at a plurality of time points at fixed time intervals in the direction opposite to the time axis with respect to the present time, and the absolute values of the plurality of sampled values taken out. The sum of the sampling values, which is the sum of the following, is calculated, and when all of these sums of sampling values exceed a predetermined threshold value, the sound starting time point is determined based on the present time point. Detection method.