JP3553828B2 - Voice storage and playback method and voice storage and playback device - Google Patents

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a voice storage / reproduction method and a voice storage / reproduction device that enable natural communication in a voice communication system or the like.
[0002]
[Prior art]
In order to transmit voice efficiently, such as when performing voice communication on a half-duplex communication path or a network, a so-called voice (voice) switch is used. When transmitting voice, open a voice line and transmit voice. There are systems in which a line is closed to allow another user to transmit voice or to transmit / receive other data. In this system, when the voice power exceeds a certain threshold by voice power measurement, the transmitting side determines the beginning of the word and opens the voice line. (Talk head) The disconnection was prevented.
[0003]
However, when the background noise is large, the speech power at the beginning of the word is lower than that of the background noise, and the speech recognition rate is low. There was a problem that it was easy to become.
[0004]
FIG. 5 shows an example of a waveform when "Kitami" is uttered, voice power, and a period in which the voice switch is ON. As is apparent from the example of FIG. 5, since the voice power does not reach the threshold at the beginning of the word "k", the ON of the voice switch is delayed, and the "k" portion of the voice at the beginning of the word is missing. You can see that As described above, the beginning of the speech is likely to occur when the speech head has a low sound power such as a consonant part, and is unlikely to occur when the speech power is high such as a vowel. Considering Japanese, speech usually has many combinations of consonants + vowels. Therefore, in order to solve the above problem, speech is always stored once, and when the beginning of a word is detected in a vowel part or the like having a large speech power, the speech is reproduced from a point in time when a certain period of time has passed, and the beginning of the speech is deleted. There are ways to prevent it.
[0005]
[Problems to be solved by the invention]
However, the accumulation of voices causes voice delays and makes the call very unnatural. Usually, when a human talks, it is considered that the utterance is often made after the speech of the other party ends. For example, considering that voice storage is performed for 100 msec, it takes 100 msec after the speaker finishes uttering until there is no line delay before the other party knows that the utterance has ended. It takes 100 msec for the user to know the utterance start of the other party. Therefore, a silent portion between conversations is 200 msec in total, and a smooth conversation is hindered.
[0006]
SUMMARY OF THE INVENTION It is an object of the present invention to prevent a disconnection of a head of speech which is a problem when realizing the above-described voice switch, and to enable a natural telephone call without causing a voice delay.
[0007]
[Means for Solving the Problems]
According to the present invention, in order to solve the above-mentioned problem, in the beginning part of the speech, the past speech is quickly reproduced from the speech accumulation unit by converting the speech speed, and the speech is input when the speech accumulation unit has no more past speech data. The audio data is reproduced at a constant speed so that no audio delay occurs at the end. According to the above-mentioned method, the silent part between conversations is at most 100 msec when the system shown above is used to store the voices and only to prevent the beginning of the conversation, and the conversations can be smoothly conducted. .
[0008]
As a result, it is possible to prevent the conversation from being unclear due to the start of the conversation, and to prevent the conversation from being unnatural due to a voice delay.
[0009]
By the way, as a device for converting the speech speed, there are devices described in JP-A-8-83095 "Method and device for speech speed conversion" and JP-A-8-202391 "Device for speech speed conversion". These are devices that output an input voice signal at a speaking speed that matches the listening ability of the listener, and mainly performs control to reduce the speaking speed. Although the speech speed varies for phonemes and speech processing frames, the speech speed is set in a large frame such as one sentence. It is not possible to play fast and eliminate the delay at the end.
[0010]
BEST MODE FOR CARRYING OUT THE INVENTION
FIG. 1 is a block diagram showing a configuration example of the present invention. In FIG. 1, reference numeral 1 denotes a voice detection unit for inputting voice and discriminating whether the input voice is a voice section, 2 a ring buffer as a voice storage unit for storing the input voice, and 3 a ring buffer 2 for input voice. A ring buffer control unit 4 for controlling the updating of the input pointer and the output pointer to store the input pointer and the output pointer, and when the voice is detected by the voice detection unit 1, determines how far back the audio is reproduced from the stored voice. The voice speed control unit 5 controls the voice speed control unit 4 to reproduce the head portion of the voice stored in the ring buffer 2 at a high speed and to reproduce the voice at a constant speed when catching up with the input voice. The speech speed converter converts the speech speed of the voice stored in the ring buffer 2.
[0011]
The sound input to the apparatus shown in FIG. 1 is measured for sound power and the like by a sound detection unit 1 and stored in a ring buffer 2. Further, the voice detection unit 1 always measures the background noise power, and dynamically changes a threshold for voice section detection.
[0012]
When the speaker does not speak, the input speech is stored in the ring buffer 2 one after another while always retaining past data. When the voice is detected by the voice detector 1, a notification is sent to the ring buffer controller 3 and the voice speed controller 4. Since the ring buffer control unit 3 knows the pointer where the audio data currently being written is stored and the pointer where the past audio data is written, the past audio data must exist. , And notifies the speech speed control unit 4 of how much past data is stored.
[0013]
The voice speed control unit 4 notifies the voice speed converter 5 of the voice speed based on the data received from the ring buffer control unit 3, reproduces all the past data within a specific time, and writes the current data. The pointer of the past input voice data is made to catch up with the pointer of the input voice data.
[0014]
For example, there is 100 msec of accumulated past data, and if one tries to catch up in 100 msec, the reproduction speed is doubled, and this information is notified to the speech speed converter 5. Conversely, it is conceivable that the speech speed notified to the speech speed converter 5 is always doubled without setting the target time.
[0015]
By the way, when a human usually adjusts the speech speed, the length of a silent portion or a vowel portion changes greatly, but the speed of a consonant portion does not change. Conversely, even if the speed of the pause portion or the vowel portion is changed without changing the speed of the consonant portion, no significant deterioration occurs in listening. In other words, since the minimum duration that can be perceived differs depending on the phoneme, when using phoneme recognition for the voice detection unit 1, it is possible to change the speech speed finely and dynamically according to the phoneme of the reproduced data, so that a call with even more uncomfortable feeling is possible. It becomes.
[0016]
Therefore, when phoneme recognition is used, information on where and how much of the stored speech data is from which phoneme is also stored, so the playback speed that guarantees the minimum duration for the phoneme in each section is set as the speech speed. The converter 5 is notified. However, if the speed is too high, the sense of incongruity increases. If the speed exceeds a predetermined maximum speed, the maximum speed is notified to the speech speed converter 5. For example, when "Ogaki" is uttered, a role of preventing "Ogaki" from being reproduced is provided.
[0017]
In addition, the minimum perceivable duration described above is small because there is little difference between certain similar phonemes, such as pauses, silence fricatives, plosives, and vowels. Instead, it is conceivable to determine the playback speed using a large classification of phonemes with a small processing amount such as a pause portion, unvoiced fricatives, plosives, vowels, and the like.
[0018]
The voice speed converter 5 extracts the audio frame data from the ring buffer 2 and compresses the frame data according to the speed specified by the voice speed control unit 4 to reduce the number of frame data. In the audio output, since the frame data is reproduced at regular intervals for each sample of the frame data, the reduction of the frame data increases the speech speed. After catching up with the input voice data, the speech speed is assumed to be the same as the input voice.
[0019]
FIG. 2 shows an example of a voice reproduced with a change in speech speed when "Kitami" is uttered. FIG. 2A shows a case where speech is detected only with power without phoneme recognition and reproduced so as to catch up with the current speech data at a constant speed. FIG. 2B shows a case where phoneme recognition is performed and phoneme recognition is performed. The state when the reproduction is performed at a variable speed that changes the reproduction speed depending on the type is shown. For convenience, a unit of audio data is called a frame. For simplicity, the speed change is expressed by thinning out the number of audio frames. “*” Represents a silent part.
[0020]
In the case of FIG. 2 (a), the reproduced sound reaches the input sound at "a" of 15 frames, and the reproduction speed is doubled until reaching the input sound. To play. In the case of FIG. 2B, since “*” in the first and second frames is a silent part preceding the plosive, the frame is reproduced at a speed skipping one frame. Since "k" of 3 to 4 frames is a consonant part, it is reproduced as it is. The “i” of 5 to 8 frames is reproduced as, for example, 3 frames in order to secure the number of frames necessary for a human to recognize a vowel. “*” Of 9 to 10 frames is 1 frame. Since “t” of the 11 to 12 frames is a consonant part, it is reproduced as it is. “A” of 13 to 17 frames is reproduced by changing 5 frames into 3 frames for the same reason as “i”. Since the past audio data will be lost thereafter, it is reproduced as it is.
[0021]
As can be seen from FIG. 2, since the input speech frame and the reproduced speech frame coincide with each other at the end, there is no speech delay at the end. When the end of the voice section is detected at the end, the ring buffer control unit 3 stops the pointer of the output voice data (hereinafter, referred to as the output pointer). When the pointer of the input voice data (hereinafter referred to as the input pointer) catches up with the output pointer, the output pointer is advanced. As a result, when an audio section is detected immediately after the end of the audio is detected and the reproduction is completed, the audio is prevented from being reproduced twice.
[0022]
FIG. 3 shows an example of control in the case where voice detection is performed only with power without phoneme recognition, and FIG. 4 shows an example of control in the case where phoneme recognition is performed and the reproduction speed is changed depending on the type of phoneme.
[0023]
When voice detection is performed only with voice (sound) power without performing phoneme recognition, first, the voice detection unit 1 detects a voice section by comparing the voice power with a threshold (S1). In the ring buffer 2, the input voice is always stored by the ring buffer control unit 3. The voice speed controller 4 receives the notification from the voice detector 1 and sets a target voice speed in a voice speed register (not shown) referred to by the voice speed converter 5 (S2).
[0024]
The voice speed converter 5 fetches voice frame data from the ring buffer 2 under the control of the voice speed controller 4 (S3), and performs voice speed conversion according to the voice speed register value (S4). The converted audio frame data is written to an output buffer (not shown) (S5), and the output pointer of the ring buffer 2 is incremented (S6).
[0025]
It is checked whether or not the input pointer of the ring buffer 2 has caught up with the output pointer (S7). If it has not caught up, the process returns to step S3, and the sound reproduction and output at the high target speech speed is repeated. If the input pointer has caught up with the output pointer, a constant speed is set in the speech speed register (S8), and the process returns to step S3 to reproduce the voice at the same speed as the input voice. The above processing is repeated until the voice section ends.
[0026]
Control for performing phoneme recognition and changing the reproduction speed according to the type of phoneme is performed as shown in FIG. In this method, a table 10 in which the minimum duration for a phoneme is stored in advance is prepared.
[0027]
First, the voice detection unit 1 performs phoneme recognition on an input voice, and detects a voice section based on the recognition result (S10). At this time, the sound power may be considered, and the detection of the sound section based on the sound power may be used together. The ring buffer 2 constantly stores input voices in the ring buffer 2, and updates the input pointer each time. The result of the phoneme recognition by the voice detection unit 1 is also stored in the ring buffer 2 together.
[0028]
When a voice section is detected, voice frame data is fetched from the ring buffer 2 (S11), and the corresponding phoneme recognition result is fetched into the speech speed control unit 4. The speech speed control unit 4 determines whether the phoneme of the voice frame data that has been processed previously is the same as the phoneme of the voice frame data that is about to be processed (S12). If they are the same, proceed to step S14. If the phoneme is a different phoneme, the process proceeds to step S13, where the duration of the phoneme is checked, the minimum duration is read from the minimum duration table 10 for the phoneme, and the speech speed determined so as not to exceed the predetermined maximum speech speed is calculated. It is set in a register (S13). Thereafter, the process proceeds to step S14.
[0029]
The voice speed converter 5 performs voice speed conversion on the voice frame data fetched from the ring buffer 2 according to the voice speed register value (S14). The converted audio frame data is written to an output buffer (not shown) (S15), and the output pointer of the ring buffer 2 is incremented (S16).
[0030]
It is checked whether or not the input pointer of the ring buffer 2 has caught up with the output pointer (S17). If it has not caught up, the process returns to step S11, and the sound reproduction output at the variable speed is similarly repeated. If the input pointer has caught up with the output pointer, a constant speed is set in the speech speed register (S18), the next voice frame data is fetched from the ring buffer 2, and the process returns to step S14 to be the same as the speed of the input voice. Play sound at speed. The above processing is repeated until the voice section ends.
[0031]
【The invention's effect】
As described above, according to the present invention, it is possible to realize a natural telephone call without causing a speech head disconnection, which is a problem when implementing a voice switch, and without causing a voice delay.
[Brief description of the drawings]
FIG. 1 is a block diagram illustrating a configuration example of the present invention.
FIG. 2 is a diagram showing a state of speech speed conversion.
FIG. 3 is a diagram illustrating a control flow in a case where voice detection is performed using only power without performing phoneme recognition.
FIG. 4 is a diagram showing a control flow when the reproduction speed is changed depending on the type of phoneme.
FIG. 5 is a diagram illustrating a relationship between a sound waveform, a sound power, and an operation of a sound switch.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 1 Voice detection part 2 Ring buffer 3 Ring buffer control part 4 Voice speed control part 5 Voice speed converter

Claims (4)

The process of inputting voice,
Determining whether the input voice is a voice section,
Storing the input voice in the voice storage means ;
When a voice section is detected, the reproduction of the stored voice is started from the position indicated by the output pointer of the voice storage means, and while the output pointer is advanced , the beginning of the stored voice is quickly reproduced, and the input voice is reproduced. The process of playing back at a constant speed after catching up ,
When the end of the voice section is detected, the output pointer is stopped until voice for a predetermined time is stored, and the output pointer is advanced when the voice is stored. Playback method. Voice detection means for inputting voice and determining whether the input voice is a voice section,
Voice storage means for storing input voice;
The storage of the input voice to the voice storage means is controlled, and when the voice section end is detected by the voice detection means, the output pointer is stopped until voice for a predetermined time is stored, and the stored voice is stopped. A voice accumulation control means for advancing an output pointer at a time ;
When a voice is detected by the voice detecting means, the reproduction of the stored voice is started from the position indicated by the output pointer, and the leading part of the voice stored in the voice storing means is quickly moved while the output pointer is advanced. Speech speed conversion control means for controlling reproduction at a constant speed when reproduced and catching up with the input voice;
A voice storage / reproduction device comprising: a voice speed conversion unit configured to convert a voice speed of voice stored in the voice storage unit under the control of the voice speed conversion control unit. Detecting the voice section based on voice power,
2. The voice storage / reproduction method according to claim 1, wherein reproduction of a speech start part of the stored voice is performed at a constant speed higher than the speed of the input voice. Performs phoneme recognition of the input speech,
2. The voice storage / reproduction method according to claim 1, wherein reproduction of a head portion of the stored voice is performed at a speed determined based on the phoneme recognition result.

Images