KR101095867B1 - Apparatus and method for producing speech - Google Patents

Abstract

According to the present invention, in the speech synthesis apparatus and method, in order to increase the intelligibility of the synthesized sound, the speech section of which the reliability is lowered according to the noise environment among the synthesized sounds generated primarily by the speech recognition unit in the speech synthesis apparatus in a noisy place or a variable noise environment. By automatically re-adjusting the parameter value to the appropriate parameter value for the noise environment, and generating the synthesized sound in the second order based on the readjusted parameter value, the synthesized sound with high clarity with respect to the noise can be obtained.

Speech Recognition, Synthesis, Noise, Clarity, Reliability

Description

Speech synthesis device and method {APPARATUS AND METHOD FOR PRODUCING SPEECH}

The present invention relates to a speech synthesis method, and in particular, in speech synthesis, a synthesized sound is readjusted to a parameter value of a voice suitable for a noise environment for a first synthesized sound in a noisy place or a variable noise environment. The present invention relates to a speech synthesis apparatus and a method for generating greater intelligibility with respect to noise.

The present invention is derived from the research conducted as part of IT new growth engine technology development project of the Ministry of Knowledge Economy and the Ministry of Information and Communication Research and Development. [Task management number: 2008-S-019-01, Task name: Portable Korean / English automatic interpretation technology Development].

Typically, text-to-speech (TTS) is a technology for converting a given text into a speech signal by speech signal processing technology, which is one of the main technical fields of speech signal processing in addition to speech recognition. As described above, the TTS records a voice signal to be reproduced in advance, such as a subway station announcement, and then replaces the voice signals with only the corresponding voice signals for some specific words. Advances have been made to advanced signal processing methods that produce natural synthesized sounds.

In addition, in recent years, a large database of voices is collected, and the voice database is segmented and labeled by phone or phone-like units, and then inputted into the input sentence. Searching for the most appropriate phonemes or strings of similar phonemes and combining them to generate synthesized sounds has been widely applied. This method can greatly improve the quality of synthesized sound. However, as the size of the voice database is very large, it is difficult to apply to a small terminal.

Recently, the Hidden Markov Model (HMM) technique, which is one of the most widely used acoustic modeling techniques, is applied to a speech model for each synthesized unit sound from a large-scale speech database. The method of learning and reproducing natural synthesized sound from this model is being actively studied.

1 illustrates a conventional speech synthesis processing flow. First, when a text sentence is input (S100), text preprocessing regarding a punctuation mark or the like is performed on the input text sentence (S102), and whether a question or a plain sentence is included. A corresponding synthesized sound control parameter is generated (S104).

Subsequently, a phonetic string conversion is performed in which an input sentence is pronounced as a phoneme string as pronounced (S106), and an HMM state string is generated (S108). In this case, the HMM status sequence includes statistical pitch information, energy, duration information, and formant information, which is a vocal tract characteristic, for a large amount of speech. Produces a synthetic sound (S110).

However, the conventional speech synthesis method as described above basically assumes a quiet use environment, and there is a problem that the intelligibility falls in a place where the noise is severe. To this end, there are ways to solve the problem by controlling the volume of the voice output device in a noisy place.However, this requires human intervention. Among them, there was a problem of judging what the sound quality is.

Accordingly, in the speech synthesis, the present invention synthesizes the synthesized sound by adjusting the parameter value of the voice suitable for the noise environment to the first synthesized sound in a place where the noise is loud or in a variable noise environment, thereby generating a synthesized sound again. It is intended to provide a speech synthesis apparatus and method that can increase the.

As described above, the present invention provides a speech synthesizer comprising: a text speech synthesizer for synthesizing an input text sentence into a speech and outputting a first synthesized sound; and a first speech recognition in a state where ambient noise is added to the first synthesized sound. The voice recognition unit detects a voice section in which the voice recognition reliability value is lower than a predetermined criterion, and readjusts the voice parameter value for the corresponding section, a voice recognized by the voice recognition unit, and a voice section in which the confidence value is low. And a voice synthesizer configured to receive the changed voice synthesis parameter value and output the second synthesized sound through the second voice synthesis.

The voice recognition unit may be configured to perform voice recognition through feature extraction on the first synthesized sound to which the ambient noise is added to obtain an HMM state string, and then use the HMM used when generating the synthesized sound before the noise is added. Computing a reliability value for each unit section of the speech recognition speech through comparison with a status string to detect a speech section whose reliability value is lower than a predetermined criterion, and analyzing the HMM status string, wherein the reliability value is lower than the predetermined criterion. The voice parameter value of the voice section is extracted and readjusted.

The present invention also provides a speech synthesis method, comprising: synthesizing an input text sentence into a voice to output a first synthesized sound, and performing voice recognition by adding ambient noise to the first synthesized sound, and then using a speech recognition reliability value. Detecting a voice section lower than the predetermined criterion, re-adjusting a voice parameter value for the detected voice section, and outputting a second synthesized sound for the voice whose voice parameter value has been adjusted.

The detecting of the speech section having a low speech recognition reliability value may include obtaining a HMM status string by performing speech recognition through feature extraction on the first synthesized sound to which the ambient noise is added. And calculating a confidence value for each unit section of the speech recognized speech by comparing the HMM state string used when generating the synthesized speech before adding, and detecting a speech section lower than the predetermined criterion.

In the present invention, in speech synthesis, in order to increase the intelligibility of the synthesized sound, the parameter value of the speech section of which the reliability is lowered according to the noise environment among the synthesized sounds generated primarily by the speech recognition unit in the speech synthesis apparatus in a place where the noise is high or in a variable noise environment. By automatically re-adjusting to a parameter value appropriate to the noise environment and generating a synthesized sound second by the readjusted parameter value, a synthesized sound with high clarity to noise can be obtained. Accordingly, there is an advantage in that it is easy to automatically generate synthesized speech having a high intelligibility in a situation where it is difficult for a person to directly adjust the intelligibility of the voice synthesis apparatus such as a subway, an airport, and a driving car.

Hereinafter, with reference to the accompanying drawings will be described in detail the operating principle of the present invention. In the following description of the present invention, if it is determined that a detailed description of a known function or configuration may unnecessarily obscure the subject matter of the present invention, the detailed description thereof will be omitted. The following terms are defined in consideration of the functions of the present invention, and may be changed according to the intentions or customs of the user, the operator, and the like. Therefore, the definition should be made based on the contents throughout the specification.

2 is a block diagram of a speech synthesis apparatus according to an embodiment of the present invention. The speech sum growth value of the present invention is a text speech synthesis unit 200, a speech recognition unit 202, a speech synthesis unit 204, A high pass filter 206.

First, the text-voice synthesizer 200 synthesizes an input text sentence into a voice and outputs a first synthesized sound. The speech recognition unit 202 performs first speech recognition in a state where ambient noise is added to the first synthesis sound output from the text speech synthesis unit 200, and then the reliability value for each unit section of the speech recognition speech is determined. The speech section detects a speech section having a reliability value lower than a predetermined reference value and readjusts the speech parameter value for the section.

The voice synthesizer 204 receives the voice recognized by the voice recognizer 202 and the changed voice parameter value of the voice section having a low reliability value and outputs the second synthesized sound through the second voice synthesis. The high frequency bandwidth filter 206 performs high frequency band filtering on the speech synthesized again through the speech synthesis unit 204.

Hereinafter, the operation of the speech synthesis apparatus will be described with reference to FIG. 2.

First, when a text sentence is input, a first synthesized sound is generated by using the text-voice synthesizer 200 with respect to the input text sentence and output to the speaker (not shown). In this case, when noise is present in the vicinity, the noise is reproduced at the voice signal level to be added to the first synthesized sound, and the added voice signal is passed through the voice recognition unit 202 to establish a predetermined standard of reliability of the first synthesized sound. A lower voice section is detected and the pitch, duration, and energy parameter values of the voice are readjusted for that section.

Subsequently, the voice signal having the readjusted voice parameter value is transmitted to the voice synthesizer 204, and the second synthesized voice is generated through the second voice synthesizer using the readjusted parameter value. Thereafter, the second synthesized sound is subjected to high frequency band filtering through the high pass filter 206 to output the final synthesized sound.

3 is a flowchart illustrating a voice synthesis operation control flow for increasing the intelligibility of synthesized sound even in a noise environment in the speech synthesis device according to an exemplary embodiment of the present invention. Hereinafter, embodiments of the present invention will be described in detail with reference to FIGS. 2 and 3.

First, when a text sentence is input (S300), the text-voice synthesizer 200 synthesizes the input text sentence into a voice and outputs a first synthesized sound (S302). As described above, the first synthesized sound output through the text-voice synthesizer 200 is added to the voice recognition unit 202 when the noise is added (S304).

Then, the voice recognition unit 202 performs the first voice recognition through feature extraction on the first synthesized sound to which the ambient noise is added (S306), and then obtains the HMM status string. Subsequently, a reliability value for each unit section of the speech recognized speech signal is calculated to detect a speech section having a reliability value lower than a predetermined reference level (S308), and for a section having a reliability value lower than a predetermined reference, the pitch and duration of the corresponding speech are continued. The parameters such as length and energy are readjusted (S310).

That is, the speech recognition unit 202 adjusts the pitch change factor value P_a that is readjusted with respect to the pitch value of the speech used when generating the first synthesized sound, and the energy change factor value that has been readjusted with respect to the energy value for a section in which the reliability value is lower than a predetermined criterion. The duration change factor value D_a readjusted with respect to the duration value used in E_a and the first synthesis sound is transmitted to the speech synthesis unit 204.

Then, the speech synthesis unit 204 generates the second synthesis sound through the second speech synthesis using the parameter value of the speech readjusted by the speech recognition unit 202 (S312).

For example, when generating the first synthesized sound, the pitch of the synthesized voice in a specific section is P (t), P (t + 1), ..., P (t + N), and the energy of the synthesized voice is E. When (t), E (t + 1), ..., E (t + N), the readjusted pitch and energy are P_a × P (t), P_a × P (t + 1),. .., P_a × P (t + N) and E_a × E (t), E_a × E (t + 1), ..., E_a × E (t + N) It can be changed in a similar manner. P_a, E_a, and D_a are real values between 0 and K, and t is a frame index of speech.

As described above, the voice parameter is reconsidered by the voice recognition unit 202 considering the surrounding noise environment, and the second synthesized sound recombined with high clarity by the voice synthesizer 204 according to the readjusted voice parameter has a high band. After the filtering for the high frequency band is performed through the pass filter unit 206 (S314), the final synthesized sound is output (S316). In this case, the reason for performing the high frequency band pass filtering is that it is known that the listening intelligibility can be further increased when the high frequency band of the voice is enhanced in the noise environment.

As described above, according to the present invention, in the speech synthesis apparatus and method, in order to increase the intelligibility of the synthesized sound, in the place where the noise is severe or in the variable noise environment, according to the noise environment among the synthesized sounds generated primarily by the speech recognition unit in the speech synthesis apparatus. By automatically re-adjusting the parameter value of the speech section with low reliability to a parameter value suitable for the noise environment, and generating the synthesized sound in the second order based on the readjusted parameter value, a synthesized sound with high clarity with respect to the noise can be obtained. In addition, when the noise environment does not change abruptly, the voice recognition is performed in subsequent speech synthesis using the same values of E_a, P_a and D_a corresponding to the energy, pitch, and duration change factor values obtained through the voice recognition process. It is possible to generate synthesized speech of high clarity without separately performing.

While the invention has been shown and described with reference to certain preferred embodiments thereof, it will be understood by those skilled in the art that various changes and modifications may be made without departing from the spirit and scope of the invention. Accordingly, the scope of the invention should not be limited by the described embodiments but should be defined by the appended claims.

1 is a flowchart of a conventional speech synthesis process;

2 is a detailed block diagram of a speech synthesis apparatus according to an embodiment of the present invention;

3 is a speech synthesis processing flow diagram that can increase the intelligibility in the noise environment according to an embodiment of the present invention.

<Brief description of the major symbols in the drawings>

200: text speech synthesis unit 202: speech recognition unit

204: voice synthesizer 206: high pass filter

Claims (10)

A text-voice synthesis unit for synthesizing the input text sentences into voices and outputting a first synthesis sound; Speech recognition unit for performing the first speech recognition in the state of adding the ambient noise to the first synthesis sound, detecting a speech section having a lower speech recognition reliability value than a predetermined reference, and readjusts the speech parameter value for the section Wow, A voice synthesizer which receives the voice recognized by the voice recognition unit and the changed voice synthesis parameter value of the voice section having a low reliability value and outputs a second synthesized sound through the second voice synthesis. Speech synthesis device comprising a. The method of claim 1, The voice recognition unit, After obtaining the HMM status string by performing voice recognition through feature extraction on the first synthesized sound to which the ambient noise is added, and comparing it with the HMM status string used when generating the synthesized sound before the noise is added, A speech synthesis apparatus for detecting a speech section having a reliability value lower than a predetermined criterion by calculating a reliability value for each section of a speech recognized speech. The method of claim 2, The voice recognition unit, And a speech parameter value of a speech section in which the reliability value is lower than the predetermined criterion through the analysis of the HMM state sequence to readjust the quality of the synthesized sound. The method of claim 3, wherein The voice parameter value is, And a parameter value for pitch, energy or duration of the speech. The method of claim 1, The apparatus comprises: A high pass filter for performing high frequency band filtering on the second synthesized speech through the speech synthesizer. Speech synthesis device further comprising. Synthesizing the input text sentence into a voice and outputting a first synthesis sound; Detecting a speech section in which the speech recognition reliability value is lower than a predetermined criterion after performing speech recognition by adding ambient noise to the first synthesized sound; Readjusting a voice parameter value for the detected voice section; Outputting a second synthesized sound for the voice whose voice parameter value is readjusted; Speech synthesis method comprising a. The method of claim 6, The detecting of the speech section having a low speech recognition reliability value may include: Obtaining a HMM status string by performing voice recognition through feature extraction on the first synthesized sound to which the ambient noise is added; Detecting a speech section lower than the predetermined criterion by calculating a reliability value for each unit section of the speech recognition speech by comparing the HMM state string used when generating the synthesized sound before the noise is added; Speech synthesis method comprising a. The method of claim 7, wherein The HMM status string is, Speech synthesis method comprising a plurality of voice parameter value information for the voice. The method of claim 8, The voice parameter value is, And a parameter value for pitch, energy, or duration of the speech. The method of claim 6, The method, Performing high frequency band filtering on the second synthesized sound after the outputting of the second synthesized sound Speech synthesis method further comprising.

Images