JPS63167400A - Voice recognition equipment for telephone - 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 Field of the Invention The present invention relates to a telephone speech recognition device that recognizes speech input from a telephone terminal.

2. Description of the Related Art In recent years, speech recognition devices for recognizing telephone voices have gradually come into general use.

A conventional telephone voice recognition device will be described below with reference to the drawings. FIG. 2 shows a conventional telephone voice recognition device. In FIG. 2, 1o and 11 are various telephone terminals, each of which is connected to a selection section 3 in a telephone exchange 2. In FIG. Further, the audio signal output from the selection section 3 is input to the analysis section e, and the acoustic feature output from the analysis section e is input to the collation section 4. On the other hand, standard pattern set 6
The output of is also connected to the matching section 4.

The operation of the telephone voice recognition device configured as described above will be described below.

The user inputs voice from the telephone terminal 1o or 11,
These voices are input to the telephone exchange 2 via the line 12 or 13. A selection unit 3 in the telephone exchange 2 determines from which telephone terminal a voice input is received, controls a switch, and inputs the voice to be recognized to an analysis unit 6. Analysis department 6
performs acoustic analysis on the input speech and outputs the extracted acoustic features to the matching section 4. On the other hand, in the standard pattern set 5, a set of statistically optimal speech analysis results for each recognition target speech is stored, and these data are also input to the collation section 4. The voice verification section receives the inputs from the two systems, performs a predetermined verification process, and outputs a recognition result. Many methods have been proposed for this matching task, but most of them are based on the concept of spectral distance between the input speech and the standard pattern. The voice is used as the recognition result of the input voice.

Problems to be Solved by the Invention However, in the above configuration, when each telephone terminal used by a user has different acoustic characteristics and different spectra are given to the voice, the spectrum performed by the matching unit 4 When comparing distances, there is a problem in that the spectral distance difference due to the difference in telephone terminals makes the difference in spectral distance due to the difference in voice itself unclear, and even leads to erroneous recognition.

As an example, FIG. 3 shows the frequency characteristics of the transmitting system of three types of commercially available telephone terminals. FIG. 3(a) shows the characteristics of a transmission system using a carbon type microphone, and the wide peak characteristics at 2 or 3 animals and the shoulder characteristics at 300 Hz are remarkable. - Figure 3 (b) shows the characteristics of a transmitting system using an electret condenser microphone, which includes the third
The remarkable properties seen in Figures (,) do not appear.

Furthermore, FIG. 3(c) shows the characteristics of another transmission system using the same electret condenser microphone, and only the wide peak characteristics at 2 or 3 appear in this.

As described above, the frequency characteristics of the transmitting system of telephone terminals differ greatly depending on the model, and it is easy to imagine that the frequency characteristics of the voice input from these terminals also differ greatly.

As is clear from the above examples, in conventional telephone voice recognition devices, in order to match each voice input from telephone terminals of different models using a single type of standard pattern, differences in the spectral distances of the voices themselves are This has the problem of being unclear and leading to misunderstandings.

In view of the above-mentioned problems, the present invention provides a telephone speech recognition device that can perform good speech recognition even when the acoustic characteristics of telephone terminals used by users are different from each other.

Means for Solving the Problems To achieve this object, the telephone voice recognition device of the present invention distinguishes and selects a plurality of telephone terminals and which telephone terminal the input voice is from. a selection section, an analysis section that acoustically analyzes the output of the selection section, a correction section that corrects the output of the analysis section, a set of standard patterns, a plurality of sets of correction coefficients to be input to the correction section, and the plurality of sets of correction coefficients to be input to the correction section. The present invention is comprised of a correspondence table that stores a predetermined many-to-one correspondence between the telephone terminal and the plurality of sets of correction coefficients, and a matching section.

Operation: With this configuration, the telephone terminal into which the speech to be recognized is input is determined, and one set of correction coefficients corresponding to the acoustic characteristics of this telephone terminal is determined. Here, each set of correction coefficients is a set of coefficients suitable for canceling out the acoustic characteristics of the telephone terminals associated with each other in the correspondence table, thereby ensuring good audio quality unaffected by differences in telephone terminals. It is possible to perform recognition.

EXAMPLE An example of the present invention will be described below with reference to the drawings.

FIG. 1 shows a telephone voice recognition device according to an embodiment of the present invention. In FIG. 1, 10 and 11 are various telephone terminals, which are connected to a selection section 3 in a telephone exchange 2 via lines 12 and 13, respectively. Further, the output selected by the selection section is input to the analysis section 6, the output of the analysis section 6 is input to the correction section 7, and the output of the correction section 7 is input to the collation section 4. On the other hand, the selection information of the selection unit 3 is input to the correspondence table 9, and the table index result output of the correspondence table 9 is input to the plurality of sets 8 of correction coefficients. Furthermore, one set of correction coefficients selected from the plurality of sets 8 of correction coefficients is input to the correction section 7. Further, the contents of the standard pattern set 6 are input to the matching section 4.

Regarding the telephone voice recognition device configured as above,
The operation will be explained below.

A user inputs voice from telephone terminal 10 or 11, and these voices are input to telephone exchange 2 via lines 12 and 13. A selection unit 3 in the telephone exchange 2 determines which line the voice input is from, controls a switch, and outputs the voice to be recognized to the analysis unit 6. For convenience of explanation, the telephone terminal used by the user for voice input will be referred to as an active terminal, and the line to which it is connected will be referred to as an active line. The selection unit 3 outputs the line number of the currently active line to the correspondence table 9.

The correspondence table 9 uses this active line number to search the many-to-10 table, determines the type of active terminal, and further calculates the acoustic characteristics of the active terminal that are mixed in the acoustic feature at the input stage of the correction unit 7. A set of correction coefficients most suitable for canceling out is selected from among the plurality of correction coefficient sets 8.

The selected set of correction coefficients is input to the correction unit 7, and the correction processing therein obtains an acoustic feature amount in which the acoustic characteristics of the active terminal are canceled.

The matching unit 4 receives this acoustic feature amount and the output from the standard pattern set 6, and performs voice matching.

Incidentally, the coefficient values in the correction coefficient set 8 and the operation contents of the correction section 7 in this embodiment must be determined in relation to the operations in the analysis section 6 and the collation section 4. -For example,
Let us consider a case where the analysis section 6 performs energy analysis by frequency band using a filter bank, and the comparison section 4 calculates the urban distance between the standard pattern and the incoming cover pattern. In this case, for example, the transmission characteristics of the transmitting system of each type of telephone terminal are measured in advance, and the obtained acoustic features are used as the set of correction coefficients.
A possible method is to subtract the correction coefficient of the telephone terminal used from the acoustic feature of the input signal in the correction section 7 during actual recognition. By this subtraction, it is possible to completely correct the transmission system of the telephone terminal under the law of addition of acoustic features. This method is suitable when an electret condenser microphone is used in the transmission system.

On the other hand, when a carbon-type microphone is used, the addition law does not hold due to its nonlinearity, and the above-mentioned correction by subtracting the acoustic feature cannot strictly hold. However, as a rough approximation, it is thought that the intended purpose can be achieved by measuring the typical frequency characteristics of the target carbon microphone and applying similar processing based on this.

As described above, according to this embodiment, a signal in which the characteristics of the telephone terminal are canceled out is always obtained as the output of the correction section 7, regardless of the telephone terminal used. Verification can be performed. Furthermore, line 1 in Figure 1
2 or 13 have specific different acoustic properties, this effect can be canceled out or reduced in a similar way.

Effects of the Invention As described above, the present invention comprises a plurality of telephone terminals, a selection section that determines and selects which telephone terminal the input voice comes from, and an analysis section that acoustically analyzes the output of the selection section. , a correction section that corrects the output of the analysis section, a set of standard patterns,
a correspondence table that stores a predetermined many-to-one correspondence between a plurality of sets of correction coefficients to be input to the correction unit, and a plurality of sets of correction coefficients and the plurality of telephone terminals;
By providing a matching section, it is possible to perform good speech recognition that cancels out differences in the acoustic characteristics of the telephone terminal models to which speech is actually input. Therefore, it is possible to realize an excellent telephone speech recognition device that can perform good speech recognition using the same method even when lines have specific and different acoustic characteristics.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of a telephone voice recognition device according to an embodiment of the present invention, FIG. 2 is a block diagram of a conventional telephone voice recognition device, and FIG. 3 a, b, and c are three types of commercially available telephones. FIG. 3 is a diagram of frequency transmission characteristics of a transmitting system of a terminal. 2... Telephone exchange, 3... Selection section, 4
....Verification section, 6.. Standard pattern set, 6.. Analysis section, 7..... Correction section, 8.
...Multiple sets of correction coefficients, 9...Correspondence table, 10.11...Telephone terminal, 12゜13
・・・・・・Line. Name of agent: Patent attorney Toshio Nakao and one other person ('
J Q yζ
＼ Figure 3 Hosu also (-) Takajishizawa (cold) Shoulder thicket (1st generation)

Claims (1)

[Claims] a plurality of telephone terminals, a selection section that determines and selects which telephone terminal the input voice comes from, an analysis section that analyzes the audio signal output of the selection section, and an acoustic feature amount of the analysis section. a correction section that corrects the output; a set of standard patterns;
a correspondence table that stores a predetermined many-to-one correspondence between a plurality of sets of correction coefficients to be input to the correction unit, and a plurality of sets of correction coefficients and the plurality of telephone terminals;
a verification unit, the selection unit determines and selects from which telephone terminal the voice to be recognized inputted from any one of the plurality of telephone terminals is inputted; - Select a corresponding set of correction coefficients from the plurality of sets of correction coefficients by referring to the correspondence table based on the selection result, and using the selected set of correction coefficients,
A speech recognition device for a telephone, characterized in that the acoustic feature output from the analysis section is corrected by the correction section, and the comparison section performs speech matching between the correction output and the set of standard patterns. .