JP7706086B2 - Voice quality evaluation device, voice quality evaluation method, and voice quality evaluation program - Google Patents

Description

The present invention relates to a technology for evaluating voice quality.

In recent years, the number of opportunities to conduct video conferences, distance learning, and other activities via communication networks has increased. This has created a need for technology that can easily evaluate the quality of this audio.

For example, Patent Document 1 proposes a speech evaluation method in which an initial speech recognition result of a sample after error correction and filtering of input speech is compared with an original text after filtering, and a speech evaluation score is calculated.
Moreover, Patent Document 2 proposes a technique in which input voice recognition data is recorded in a storage unit, and a search unit matches the voice recognition data with dictionary data to generate a voice recognition result.

JP 2016-51179 A JP 2018-54717 A

A subjective quality assessment method called MOS (Mean Opinion Score) has long been used to assess the voice quality of telephone systems. There is also an objective quality method called POLQA (Perceptual Objective Listening Quality Assessment) that uses a computer to estimate the results of subjective assessments using MOS. POLQA is a method that calculates a MOS value by comparing an original voice, called the reference voice, with a voice recorded on the receiving side.

When evaluating the quality of audio in video conferences, distance learning, etc., it is essential that the receiver can accurately hear what the speaker is saying for a period ranging from several tens of minutes to several hours. However, POLQA is not suitable for evaluating the quality of audio that is transmitted over long periods of time. POLQA sets standards for reference audio and reference video, and does not target evaluation of normal conversation or audio that does not meet the standards. In addition, since POLQA is a method of comparing audio data, the amount of data becomes large when attempting to evaluate long audio data, and the processing costs are enormous. For this reason, it is difficult to evaluate changes over time, such as "how packetized audio and video are affected by changes in network conditions."

Furthermore, because POLQA is a method for evaluating the difference between the original voice and the voice of the receiver, it cannot be said to adequately evaluate the accuracy of the speech content that can be heard by the receiver. For example, even if equalization or other processes to improve the voice quality have been performed, POLQA will determine that there is a difference in the voice and give a low rating. Similarly, in situations where the collected sound data contains spatial noise or reverberation, such as video conferences or distance learning, even if the data is played back after processing to remove the noise, it will still be determined that there is a difference in the voice and will be given a low rating.

The present invention aims to easily evaluate the transmission quality of voice data transmitted over a network.

To achieve the above object, a voice quality evaluation device according to one aspect of the present invention comprises a voice generation unit that converts reference text into reference voice data, a reference voice transmission unit that transmits the reference voice data generated by the voice generation unit to a playback device via a network, an evaluation target voice acquisition unit that acquires evaluation target voice data played back from the playback device, a voice recognition unit that performs voice recognition on words included in the evaluation target voice data to generate evaluation target text, and an evaluation unit that performs quality evaluation of the voice received via the network based on the evaluation target text.

The evaluation unit may compare the evaluation target text with the reference text to perform the quality evaluation.

The speech recognition unit may perform speech recognition on words contained in the reference speech data to generate a second reference text, and the evaluation unit may perform the quality evaluation by comparing the evaluation target text with the second reference text.

The speech recognition unit may perform speech recognition of words contained in the reference speech data to generate a second reference text, and the evaluation unit may perform a first evaluation by comparing the reference text with the second reference text, and a second evaluation by comparing the reference text with the text to be evaluated, and perform the quality evaluation based on the results of the first evaluation and the second evaluation.

To achieve the above object, a voice quality evaluation method according to another aspect of the present invention includes a voice generation process for converting reference text into reference voice data, a reference voice transmission process for transmitting the reference voice data generated by the voice generation process to a playback device via a network, an evaluation target voice acquisition process for acquiring evaluation target voice data played back from the playback device, a voice recognition process for voice-recognizing words included in the evaluation target voice data to generate evaluation target text, and an evaluation process for evaluating the quality of the voice received via the network based on the evaluation target text.

In order to achieve the above-mentioned object, a voice quality evaluation program according to yet another aspect of the present invention causes a computer to execute a voice generation instruction for converting reference text into reference voice data, a reference voice transmission instruction for transmitting the reference voice data generated by the voice generation instruction to a playback device via a network, an evaluation target voice acquisition instruction for acquiring evaluation target voice data played back from the playback device, a voice recognition instruction for voice-recognizing words contained in the evaluation target voice data to generate evaluation target text, and an evaluation instruction for evaluating the quality of the voice received via the network based on the evaluation target text.
The computer program may be provided by being stored on various data-readable recording media, or may be provided so as to be downloadable via a network such as the Internet.

The voice quality evaluation device according to the present invention makes it possible to easily evaluate the transmission quality of voice data transmitted over a network.

1 is a functional block diagram of a sound quality evaluation device according to a first embodiment of the present invention; FIG. 1 is a conceptual diagram showing an example of data processed in the voice quality evaluation device, including (a) an example of reference text, (b) an example of reference voice data reading the reference text, (c) an example of voice data to be evaluated received via a network, and (d) an example of text to be evaluated obtained by performing voice recognition on the voice data received via the network. 3 is a flow chart showing the flow of a series of processes executed by the sound quality evaluation device. 10 is a flow chart showing the flow of a series of processes executed by a sound quality evaluation device according to a second embodiment of the present invention. 13 is a flow chart showing the flow of a series of processes executed by a sound quality evaluation device according to a third embodiment of the present invention.

Below, a voice quality evaluation device, a voice quality evaluation method, and a voice quality evaluation program according to an embodiment of the present invention will be described with reference to the drawings.

First Embodiment
Configuration of the Sound Quality Evaluation Device As shown in Figure 1, the sound quality evaluation device 1 is a device for evaluating the quality of sound transmitted and received via a network NW. The network NW may be, for example, the Internet or any other appropriate communication line connected by wire or wirelessly, and may take any form.

For example, the sound quality evaluation device 1 is connected to a playback device 50 via a network NW. The playback device 50 is a terminal such as a personal computer, a smartphone, or a tablet, and is the terminal on which the viewer of the video call is watching.

The sound quality evaluation device 1 transmits sound data of reference sound (hereinafter also referred to as "reference sound data") to the playback device 50 via the network NW. The playback device 50 converts this reference sound data into sound and plays it back. Note that the reference sound data may be received by the playback device 50 via an appropriate device on the network NW, for example a server. The sound quality evaluation device 1 acquires the sound data transmitted via the network NW and played back by the playback device 50, and evaluates the quality of this sound.

The voice quality evaluation device 1 is composed of a storage medium such as a memory, a processor, a communication module, and an input/output interface, and the processor executes a computer program recorded in the storage medium to realize the functional blocks shown in FIG. 1. The storage medium is a computer-readable recording medium, and may include storage devices such as RAM (random access memory), ROM (read only memory), disk drives, SSDs (solid state drives), and flash memories. Here, non-transient storage devices such as ROM, disk drives, SSDs, and flash memories may be included in the voice quality evaluation device 1 as separate storage devices distinct from the memory.

The voice quality evaluation device 1 has the above-mentioned hardware configuration and mainly comprises, for example, a reference text acquisition unit 11, a voice generation unit 12, a reference voice transmission unit 13, an evaluation target voice acquisition unit 14, a voice recognition unit 15, and an evaluation unit 16. Note that part or all of the configuration of the voice quality evaluation device 1 may be realized by another hardware configuration, or part or all of it may be realized by a cloud computer. Furthermore, part of the functions of the voice quality evaluation device 1 may be configured inside the playback device 50. In this case, for example, the evaluation target voice acquisition unit 14, the voice recognition unit 15, and the evaluation unit 16 may be configured in the playback device 50.

The reference text acquisition unit 11 is a functional unit that acquires reference text to be used in the speech quality evaluation. The reference text is, for example, text data as shown in FIG. 2(a) and may be in any suitable language other than Japanese. The reference text acquisition unit 11 may acquire the reference text via a suitable network, or may accept input via suitable input means possessed by the speech quality evaluation device 1.

As shown in the conceptual diagram of FIG. 2(b), the voice generation unit 12 is a functional unit that converts the reference text acquired by the reference text acquisition unit 11 into reference voice data. The voice generation unit 12 may read the reference text using an artificial voice and convert it into reference voice data. The voice generation unit 12 may also be configured to display the reference text on a display unit such as a display, and prompt a speaker who can speak accurately, such as an announcer, to read it aloud. In this case, the voice generation unit 12 is configured to pick up the voice read by the speaker.

The reference audio transmission unit 13 is a functional unit that transmits the reference audio data generated by the audio generation unit 12 to the playback device 50 via the network NW. The playback device 50 plays the received reference audio data as audio. At this time, the playback device 50 may perform appropriate signal processing on the audio data before playing it back. This signal processing may be, for example, acoustic signal processing such as equalization processing, noise canceling processing, frequency filtering processing, or amplification processing, to make the words included in the audio more clearly audible, or may be processing to complement information lost due to transmission over the network NW.

The reference audio data may be received by the playback device 50 via an appropriate device on the network NW, such as a server. The above-mentioned signal processing may be performed on the reference audio data by the device.

The evaluation target voice acquisition unit 14 is a functional unit that acquires voice data (hereinafter, also referred to as "evaluation target voice data") played back from the playback device 50. The evaluation target voice acquisition unit 14 is connected to the playback device 50 and acquires the voice data.
As shown in the conceptual diagram of Fig. 2(c), the evaluation target voice data is partially different from the reference voice data. In the example of the figure, the amplitude of some of the data shown in area L is reduced. The evaluation target voice data may be data that has been degraded from the reference voice data during transmission, or may be data that has been processed to make the speech easier to hear by the appropriate signal processing described above performed in the playback device 50 or a device on the network NW.

The speech recognition unit 15 is a functional unit that performs speech recognition on words contained in the speech data to be evaluated and generates a text to be evaluated. The example in FIG. 2(d) shows that a portion of the text shown in area W differs from the reference text.

The evaluation unit 16 is a functional unit that performs quality evaluation of the speech received via the network NW based on the evaluation target text. In this embodiment, the evaluation unit 16 performs quality evaluation of the speech by comparing the evaluation target text with a reference text.
The voice quality evaluation score is calculated, for example, by the following formula (1).

Quality evaluation score = number of correctly recognized syllables in the evaluation target text / number of syllables in the reference text x 100
...(1)

The higher the quality evaluation score, the better the voice quality. If all sounds are accurately recognized in the text to be evaluated, the quality evaluation score will be 100. In the example of FIG. 2, the reference text has 93 sounds, of which 89 have been accurately recognized in the text to be evaluated, resulting in a quality evaluation score of 96. The quality evaluation score is displayed on an appropriate display unit possessed by or connected to the voice quality evaluation device 1.

The evaluation unit 16 may calculate one quality evaluation score for the entire text to be evaluated, or may divide the text to be evaluated into multiple parts and calculate a quality evaluation score for each part, thereby calculating multiple quality evaluation scores along the time axis for one reference text. A configuration in which multiple quality evaluation scores are calculated for one reference text makes it possible to evaluate changes over time in the transmission state of the network. In this case, the evaluation unit 16 may divide the text to be evaluated so that there is no overlap between the parts, or may divide the text to be evaluated while allowing some overlap on the time axis. The voice recognition unit 15 may divide the voice data and then convert each part into text data. In this case, the evaluation unit 16 performs an evaluation on each piece of text data.

Processing Flow The processing flow in which the sound quality evaluation unit 1 evaluates transmitted sounds will now be described with reference to FIG.
First, the reference text acquisition unit 11 acquires the reference text (step S11). Next, the voice generation unit 12 acquires reference voice data based on the reference text (step S12). Next, the reference voice transmission unit 13 transmits the reference voice to the playback device 50 (step S13).

Then, the evaluation target voice acquisition unit 14 acquires the target voice data played on the playback device 50 (step S14). The voice recognition unit 15 then performs voice recognition on the target voice data to generate text data (step S15). The evaluation unit 16 evaluates the matching rate between the reference text and the evaluation target text and calculates a quality evaluation score (step S16). The quality evaluation score is then displayed on an appropriate display unit (step S17).

The voice quality evaluation device according to the present invention can evaluate the quality of transmission over a network by focusing on whether the speaker's speech content can be accurately heard on the receiving side. Furthermore, the voice quality evaluation device according to the present invention compares and evaluates text data, so the amount of data to be analyzed can be compressed compared to a configuration in which voice data is compared. This makes it possible to evaluate the quality of transmission for long periods of speech. Furthermore, a configuration in which the speech to be evaluated is converted into text data and evaluated can make it possible to appropriately evaluate whether the speech content is being transmitted accurately, even when processing to improve sound quality has been performed.

Second Embodiment
A voice quality evaluation device according to a second embodiment of the present invention will be described, focusing on the differences from the first embodiment. In this embodiment, a voice recognition unit 15 performs voice recognition on words contained in the reference voice data to generate a second reference text, and an evaluation unit 16 compares the evaluation target text with the second reference text to evaluate the quality of the voice. Note that explanations of configurations similar to those in the first embodiment will be omitted where appropriate, and the same reference symbols will be used.

As shown in FIG. 4, in the voice quality evaluation device according to the second embodiment, a reference text is acquired (step S11), reference voice data is generated (step S12), and then the voice recognition unit 15 performs voice recognition on the reference voice data to generate a second reference text (step S21). In addition, the reference voice data is transmitted (step S13), evaluation target voice data is acquired via the playback device 50 (step S14), and voice recognition is performed (step S15). The order of step S21 and steps S13 to S15 is arbitrary, and they may be performed simultaneously.

Next, the evaluation unit 16 evaluates the matching rate between the second reference text and the text to be evaluated, calculates a quality evaluation score (step S22), and displays this quality evaluation score on the display unit (step S23).
In this case, the quality evaluation score is expressed, for example, by the following formula (2).

Quality evaluation score = number of correctly recognized syllables in the evaluation target text / number of syllables in the second reference text × 100
... (2)

According to this configuration, if there is a recognition error by the speech recognition unit 15, the same recognition error will appear in both the second reference text and the text to be evaluated, so the effect of the recognition error on the quality evaluation score can be removed. In other words, in this embodiment, the speech quality can be evaluated excluding the effect of the speech recognition unit 15.

Third Embodiment
A voice quality evaluation device according to a third embodiment of the present invention will be described below, focusing on the differences from the second embodiment. In this embodiment, the voice recognition unit 15 performs voice recognition on words contained in the reference voice data to generate a second reference text, and the evaluation unit 16 performs a first evaluation by comparing the reference text with the second reference text and a second evaluation by comparing the reference text with the text to be evaluated, and then performs a voice quality evaluation based on the results of the first evaluation and the second evaluation.
In addition, explanations of configurations similar to those of the first or second embodiment will be omitted as appropriate, and the same reference numerals will be used.

As shown in Fig. 5, in the voice quality evaluation device according to the third embodiment, a reference text is acquired (step S11), reference voice data is generated (step S12), and then the voice recognition unit 15 performs voice recognition on the reference voice data to generate a second reference text (step S21). Next, a first evaluation is performed to calculate the matching rate between the reference text and the second reference text (hereinafter also referred to as the "first matching rate") (step S31).

Furthermore, reference voice data is transmitted (step S13), and the voice data to be evaluated is obtained via the playback device 50 (step S14), and voice recognition is performed (step S15). Next, a second evaluation is performed (step S32) to calculate the matching rate between the reference text and the text to be evaluated (hereinafter also referred to as the "second matching rate"). The order of steps S21 and S31 and steps S13 to S15 and step S32 is arbitrary, and they may be performed simultaneously.

Next, the evaluation unit 16 compares the first and second match rates to calculate a quality evaluation score (step S33), and displays the quality evaluation score on the display unit (step S34). For example, the quality evaluation score is expressed by the following formula (3).

Quality evaluation score = second match rate / first match rate × 100 (3)

In addition to the quality evaluation score, the first match rate and the second match rate may be displayed on the display unit.

With this configuration, the accuracy of the voice recognition by the voice recognition unit 15 can be confirmed using the first and second match rates, and the quality evaluation score can be used to confirm the evaluation of the voice quality before and after transmission.

In this way, the voice quality evaluation device according to the present invention makes it possible to easily evaluate the transmission quality of voice data transmitted over a network.

In this explanation, the evaluation of voice quality before and after transmission over a network has been taken as an example, but the voice quality evaluation device according to the present invention can be used to evaluate not only networks but also any mechanism for transmitting spoken content. Furthermore, the voice quality evaluation device can be used not only to evaluate networks, but also to help with accurate information sharing in video conferences, etc., by building a system that allows speakers in video conferences or remote classes to visually check parts that were not transmitted accurately in near real time or after the fact, and prompts them to speak those parts again.

Reference Signs List 1 Voice quality evaluation device 11 Reference text acquisition unit 12 Voice generation unit 13 Reference voice transmission unit 14 Evaluation target voice acquisition unit 15 Voice recognition unit 16 Evaluation unit NW Network

Claims (6)

a speech generator for converting the reference text into reference speech data;
a reference voice transmission unit that transmits the reference voice data generated by the voice generation unit to a playback device via a network;
an evaluation target voice acquisition unit that acquires evaluation target voice data reproduced from the reproduction device;
a speech recognition unit that recognizes words included in the evaluation target speech data and generates an evaluation target text;
an evaluation unit that evaluates the quality of the voice received via the network based on the evaluation target text;
Equipped with
The speech recognition unit generates a second reference text by performing speech recognition on words included in the reference speech data;
the evaluation unit compares the evaluation target text with the second reference text to perform the quality evaluation.
Voice quality assessment device.
a speech generator for converting the reference text into reference speech data;
a reference voice transmission unit that transmits the reference voice data generated by the voice generation unit to a playback device via a network;
an evaluation target voice acquisition unit that acquires evaluation target voice data reproduced from the reproduction device;
a speech recognition unit that recognizes words included in the evaluation target speech data and generates an evaluation target text;
an evaluation unit that evaluates the quality of the voice received via the network based on the evaluation target text;
Equipped with
The speech recognition unit generates a second reference text by performing speech recognition on words included in the reference speech data;
the evaluation unit performs a first evaluation by comparing the reference text with the second reference text, performs a second evaluation by comparing the reference text with the evaluation target text, and performs the quality evaluation based on results of the first evaluation and the second evaluation.
Voice quality assessment device.
a speech generation process for converting the reference text into reference speech data;
a reference voice transmission process for transmitting the reference voice data generated by the voice generation process to a playback device via a network;
an evaluation target voice acquisition process for acquiring evaluation target voice data reproduced from the reproduction device;
A speech recognition process for generating an evaluation target text by performing speech recognition on words included in the evaluation target speech data;
an evaluation process for evaluating the quality of the voice received via the network based on the evaluation target text;
Including,
The speech recognition process includes performing speech recognition on words included in the reference speech data to generate a second reference text;
the evaluation process compares the evaluation target text with the second reference text to perform the quality evaluation;
Voice quality assessment methods.
a speech generation process for converting the reference text into reference speech data;
a reference voice transmission process for transmitting the reference voice data generated by the voice generation process to a playback device via a network;
an evaluation target voice acquisition process for acquiring evaluation target voice data reproduced from the reproduction device;
A speech recognition process for generating an evaluation target text by performing speech recognition on words included in the evaluation target speech data;
an evaluation process for evaluating the quality of the voice received via the network based on the evaluation target text;
Including,
The speech recognition process includes performing speech recognition on words included in the reference speech data to generate a second reference text;
the evaluation process includes comparing the reference text with the second reference text to perform a first evaluation, comparing the reference text with the evaluation target text to perform a second evaluation, and performing the quality evaluation based on results of the first evaluation and the second evaluation.
Voice quality assessment methods.
speech generation instructions for converting the reference text into reference speech data;
a reference audio transmission command for transmitting the reference audio data generated by the audio generation command to a playback device via a network;
an evaluation target voice acquisition command for acquiring evaluation target voice data reproduced from the reproduction device;
A voice recognition command for generating a text to be evaluated by performing voice recognition on words included in the voice data to be evaluated;
an evaluation command for evaluating the quality of the voice received via the network based on the evaluation target text;
on the computer ,
The speech recognition command includes performing speech recognition on words included in the reference speech data to generate a second reference text;
the evaluation instruction compares the evaluation target text with the second reference text to perform the quality evaluation;
Voice quality evaluation program.
speech generation instructions for converting the reference text into reference speech data;
a reference audio transmission command for transmitting the reference audio data generated by the audio generation command to a playback device via a network;
an evaluation target voice acquisition command for acquiring evaluation target voice data reproduced from the reproduction device;
A voice recognition command for generating a text to be evaluated by performing voice recognition on words included in the voice data to be evaluated;
an evaluation command for evaluating the quality of the voice received via the network based on the evaluation target text;
Run the following on your computer:
The speech recognition command includes performing speech recognition on words included in the reference speech data to generate a second reference text;
the evaluation instruction includes comparing the reference text with the second reference text to perform a first evaluation, comparing the reference text with the evaluation target text to perform a second evaluation, and performing the quality evaluation based on results of the first evaluation and the second evaluation.
Voice quality evaluation program.

Images