TW535140B - Phoneme decoder - Google Patents

Abstract

The present invention provides a phoneme decoder, which uses the voice data encoded with three parameters in the voice data, such as pitch, root of mean square, and reflection of coefficients, for being composed as a voice signal. The phoneme decoder comprises: an initialization unit, a parameter loading unit, a smoothing unit, a composition unit, and a voice output unit. The initialization unit generates the initialization signal, and loads the three voice parameters, such as pitch, root of mean square, and reflection coefficients, into the smoothing unit. The smoothing unit receives the voice parameter data, and uses the interpolation method for smoothing, and transmits the processed voice parameter data to the composition unit. The composition unit composes the voice data according to the parametric sequence of pitch, reflection coefficients and root of mean square, and outputs the same to the voice output unit.

Description

535140 V. Description of the invention d) [Application field of the invention] The present invention is about a language basis to encode the speech [Background of the invention] In the low-to-medium-level electronic dictionary, 'the electronic dictionary can become the main sub-dictionary in the market competition Power, improvement, and at the same time can reduce the recording of specific speech, due to its large limitations, it is quite costly, the synthesis method to approach real human hair, 5 voice, body, and the sound quality of this voice analysis and synthesis of speech analysis The number of models generated by the proposed speech must be synthesized, and the corresponding speech coding represents the original signal due to the data synthesized by speech analysis. Therefore, it involves the adaptive difference in the sampling waveform encoding of speech. Pulse Code Modulation As for the reconstructed signal and the original letter, it uses the least mean square error synthesizer, especially a speech phoneme decoder for decoding. The ‘flag’ in the market is characterized by a real human voice. In order to improve, all manufacturers have focused on voice production costs. Some manufacturers emphasize the large amount of data and the output of the system. Therefore, most manufacturers use voice to save electronic dictionaries. The technology is to use these voices in accordance with certain processing parameters. Therefore, according to different methods and speech synthesis methods. The process is to minimize the sound signal, which is also commonly referred to as speech compression and encoding and decoding technologies. Pulse code modulation (Adaptive m, A D P C M) encoding method, the better the Lu- number waveform, the better; from the mathematical difference criterion (Minimum Mean

Error Criterion), but the bit rate of the ADPCM method is lower than that of the phoneme for low-order electrical functions. Real humans receive extreme sound analysis voice data method. Perspective Square

535140 V. Description of the Invention (2) 2 4kbps (Kilo Bit Per Second), there will be problems that the quality of the sound after the reduction is poor and the amount of calculation is large. The above-mentioned speech analysis and synthesis has the advantages of greatly compressing the amount of speech data, and also has the advantage of confidential communication (using encryption technology). However, its shortcomings are the importance of speech synthesis, partial separation, and base period, which are often different from natural speech, resulting in unnatural and even difficult to identify shortcomings.

Even with compressed speech analysis and synthesis technology, there is still the possibility of saving memory space. In addition, the existing speech analysis and synthesis technologies mostly operate on-line. Therefore, it is necessary to add the action of judging whether the voice has "sound". Often, in the process of judging, "sound" and "sound" are used. The "soundless" part of the judgment is incorrect, which leads to a hoarse sound during speech synthesis. Therefore, how can the speech generated by the speech analysis and synthesis technology be close to natural speech, that is, the improvement of the sound quality; on the other hand, how can the maximum compression be achieved, that is, the least memory consumption? Space; on the other hand, how to make the process of speech analysis and synthesis relatively simple; the above points have become important research topics.

[Objective and Summary of the Invention] In view of the problems of the conventional technology above, the present invention provides a coding method based on speech phoneme classification, which uses speech phonemes to be classified into voiced, silent, and mute, as long as the voiced part is encoded; decoded At this time, as long as the speech phoneme decoder of the present invention is applied to the voiced encoding part, a high-computation speech decoding can be performed.

Page 5 535140 V. Description of the invention (3) According to the technology disclosed by the present invention, the present invention provides a speech phoneme decoder, which uses an amplitude parameter (RMS), a base period parameter (Pi tch) and a linear prediction. Estimating the encoding method (Linear Prective Coding, LPC) to decode the speech data of the side code of the spectral parameter (RC's), the encoded speech data is stored in a speech database, and the speech data database The speech data of the present invention is decoded. The speech phoneme decoder of the present invention includes: an initialization unit, a loading parameter unit, a smoothing processing unit, a synthesis unit, and a speech output unit.

The initialization unit is used to generate an initialization signal (i n i t i a 1). The loading parameter unit is connected to the initialization unit, and is used to receive the initialization signal sent by the initialization unit, and load the voice data of the voice frame from the voice database by using a voice frame (F r a m e) as a unit. The smoothing processing unit is used to receive the voice data of the sound frame loaded by the loading parameter unit, and uses a base period (P itch) of the sound frame as the length to process the sound separately using the internal difference method. The amplitude parameter, base period parameter and spectrum parameter in the voice data of the frame. After the smoothing processing unit processes the voice data of the sound frame, it sends a signal of the next sound frame to the loading parameter unit to load the next sound frame. Voice data. The synthesis unit is used to receive the voice data of each base cycle processed by the smoothing processing unit and synthesize it into a voice signal; after the synthesis unit has processed the voice data of each base cycle, it sends a signal of the next base cycle to the smoothing The processing unit processes the voice data of the next base week. Finally, the synthesis unit can send the synthesized voice signal to the voice output unit to output the voice. In addition, smoothing is handled by the internal difference method, which must be calculated

Page 6 535140 V. Description of the invention (4) Proportion parameter (Prop; Proportion). In addition, because the base unit is used as the synthesis unit during synthesis, that is, one cycle at a time. Therefore, the total length of the period synthesized in a sound frame must be less than the length of the speech to be synthesized (Frame_len), and the remaining unsynthesized speech length (Frame —res = Frame —len — Synths) ^ 1 and only until the next sound Marina processing, so the speech length to be synthesized in the next sound box is

Frame_1en = Frame_res + 180〇 where Pr〇p = (Synths + Pitchl) / Frame_leno

With regard to the features and implementation of the present invention, the preferred embodiment is described in detail with reference to the drawings: [Detailed description of the invention]

Because the electronic dictionary market has relatively regular speech processing and requires a large amount of data compression, the present invention uses the Linear Predictive Coding (hereinafter referred to as LPC) method as the encoding and decoding method of the present invention. Because this method is based on a speech utterance model, the channel filter (V 0ca 1 Tract F i 11 er) of the estimated signal and the basic period (P it ch) are used to achieve the purpose of compression, which can reach very low bits. The element rate (L ow B it Rate) is quite suitable as the encoding method of the present invention. The present invention uses the "voiced", "unvoiced" and "silent" voice phonemes to classify basic sounds, and compresses and encodes the voiced phoneme parts, and the voiced phoneme parts The original sound is left uncompressed, and the mute part only records the mute length. The parameters calculated in this classification include the amplitude (R M S;

Page 7 535140 V. Description of the invention (5) There are three types of parameters: root of mean square), pitch (ie pitch) and spectrum (RC's; reflection coefficients) parameters. Among them, the acquisition of the amplitude parameter and the base period parameter is based on a sound frame (a sound frame: 180 sampling points, a sampling rate of 8 kHz) as a unit, and the parameter values are calculated step by step. The spectrum parameters (RC, s) are obtained according to the LpC model, that is, according to the following transfer function (T r a n s 丨 e r

Function in Z-Domain) H (z) is calculated as: H (z) two A0 / (l + alz-l + a2z-2 ·· + ai〇z—i〇) where AO is the amplitude parameter, z (= e_jwM, is a complex number (c0_iex number), al ~ alO are LPC parameters. From the above three parameters, a "voice" voice sound box (180sampleS) can be encoded into 54 bits, compressed bit rate Equivalent to 2.4 kbps, the bit configuration of each parameter is as follows: 6 I 6 5 ~ 5 ^ 1 ^^ 1 5 4 4 4 4 3 3 The encoded speech obtained by the θ 9 prime encoding method needs to be decompressed Use the internal difference method to smooth the amplitude, base period, and frequency of the voiced speech part! Then use the speech synthesizer to restore the voiced speech, “... the ear mouth F knife /, and take out the original voice according to the address For the mute part, just take out the long silence time. The voice database established by the above method recognizes the above three parameters as the basis for encoding.

Page 8 535140 5 Description of the invention (6) The rules can be used to design speech phoneme decoder.

Yuan Jing and Pi Pi speech phoneme decoder actions have been performed. . ", Sequence (Bi t Stream), that is," ^ "First, a series of bit-coded speech data is first converted into the selected spectral parameters in the speech database, and then these parameters are faced with: , The amplitude and pitch are synthesized with a base period (Pi tcW is a single-two δ ^ synthesizer). Combine a set of parameters and store the previous frame \ mother frame a frame read; f ru η \ Bu g frame number (RMS0, RC0 ltCh〇), a period synthesis of the mother, team U, itch), from these sound frames using a \ number of wheat (RMS, KC, (s 嶋 ther) to get. For smoothing, smoothing uses the internal difference method to make a sound ^ (Prop;, .. ^ Η, Λ force ^ because the base period is used to synthesize early in the synthesis, that is, one cycle is synthesized at a time. Therefore, in _ Sound frame = total length of two cycles' must be less than the length of the speech to be synthesized, (Frame — len), and the remaining unsynthesized speech length (Frame_res = Fraine — len-Synths) will be processed in the next sound frame, so The speech length to be synthesized in the next frame is

Frame_len = Frarae_res + 180. Among them, Prop: (synths:) itchl) / Frame_len0, the speech phoneme decoder designed by the present invention using the above-mentioned speech coding method will be described in detail. & First, please refer to "Figure 1", the system architecture diagram of the speech phoneme decoder of the present invention, which includes the following parts: Initialization unit 1 0, loading parameter unit 20, smoothing processing unit 30, synthesis Unit 40 and voice input 535140 V. Description of the invention (7) Output unit 50. First, the initialization unit 10 generates an initialization signal (i n i t i a 1), and loads the parameter unit 20 to set the initial values of the parameters accordingly. Then, the parameter loading unit 20 sequentially loads all parameter values in the frame to be synthesized, that is, loads three speech parameters in one frame at a time. Next, the smoothing processing unit 30 smoothes various speech parameters loaded in the loading parameter unit 20, processes one pitch (one pitch) in one sound frame at a time, and sends these smoothed parameters The synthesis unit 40 synthesizes the speech, and sends a signal of "Next sound frame n (Next_Frame)" to the loading parameter unit 20, and it loads the speech parameters of the next sound frame ". The synthesis unit 40 The synthesized voice signal is sent to the voice output unit 50 to output the voice, and the "Next_Pi tch" signal is sent to the smoothing unit 30, which allows the smoothing unit to process " the voice of the next base cycle π Parameters. Next, the speech phoneme decoder of the present invention will be described with specific embodiments, please continue to refer to "Figure 2", which illustrates the signal transmission architecture of the present invention. The initialization unit 10 generates an initialization signal (i n i t i a 1). Load parameter unit 20 sets the initial values according to the initialization signal. In addition, it is responsible for loading the three parameters of speech phonemes (RCj (10), RMSj, Pitch]), and retains the three parameters of the previous sound box. (RCO (IO), RMSO, Pitch 0). Finally, according to the synthesized length (L) of each frame processing sent by the smoothing unit to generate the next frame processing length (M). The smooth processing unit 30 receives the parameters transmitted by the loading parameter unit 20, and then changes the three parameters of the sound box (RC j (10), RMSj,

Page 10 535140 V. Description of the invention (8)

Pi tch j) is smoothed, and the processed parameters (RC (1 0), RMS, Pitch) are transmitted to the synthesis unit 40 in units of one base cycle (Pitch j) at a time, and the next base cycle is sent out The (Next_Pitch) signal requires the smoothing processing unit 30 to transmit the parameters of the next base cycle, and sends a synthesis length (L) to the loading parameter unit 20, that is, the length of the synthesized speech this time. Finally, the synthesis unit 40 synthesizes the three parameters and sends them to the voice output unit 50 to output the voice. Among them, the bit signal of each parameter, as shown in "Figure 2", the initialization signal (initial) is a one-bit control signal; RC0 (10) is an eight-bit signal with a sign (signed); RMS 0 Is an unsigned sixteen-bit signal; P itch 0 is an unsigned eight-bit signal; RC j (1 0) is a signed eight-bit signal; RMS j is an unsigned ten Six-bit signal; P it ch j is an unsigned eight-bit signal; the length of the synthesized sound frame M is an unsigned nine-bit signal; the synthesized length L is an unsigned nine-bit signal; RC is a signed Signed eight-bit signal; RMS is unsigned sixteen-bit signal; Pitch is unsigned eight-bit signal; N e X t _ Frame is a one-bit control signal; N e X t _ Pitch is a one-bit control signal; the sixteen-bit signal with a sign sent by the synthesis unit. Then please refer to "Figure 3", which illustrates the signal generation architecture diagram from the initialization unit to the parameter loading unit. First, the initialization signal ninitia Γ generated by the initialization unit 10 allows the loading parameter unit 20 to set initial values of various parameters, including the synthesized length (L = 0) and the length of the synthesized sound frame (M 2 180 samples). Points, taking the sampling rate of 8 0 0 times per second as an example), amplitude (RMSO-20) (RMSj; Root of Mean Square), pitch

535140 V. Description of the invention (9)

(PitchO ^ Pitchl) (Pitchj; basic period of the j-th frame > and spectrum parameters (RCO (i) = RCl (i), i 2 0, 1, 2, ..., 9; Renecti Coefficients), etc. Etc. The operation of reading data is performed by the data loading unit 24, wherein the bit sequence (54 bits) is the phoneme data of the voice input from the voice database, and is decoded into rCKi by the parameter decoder 241. ): RMs · Pi tch j and other parts are input into the second register 25 respectively. + J The second temporary storage is 2 5 to transfer the read data to the next neighboring eight = that is, the smoothing processing unit 3 0 and the third temporary storage device 2 6. Sang: The data read by Er Yi this time is temporarily stored, and can be used as the next: 9 ^ Device 26 will refer to the sound parameter reference data, that is, when the second king of Ping Yan is received ;: a sound box of the read language (Next—Frame) command, that is, two or two; the parameter value of the previous sound frame (RC0 (1〇), RMs〇_ — the number of people is set to the smoothing unit 3Q for smoothing processing (Refer to 1 tCh0), and enter, bu, due to the initial frame length (18〇 ^^ ^ the length of the base cycle, there will be a remaining part. Hand full double the minutes into the next frame length In this case, the remaining part (inUial) is rotated into the temporary register and the calculation f is initialized first, the temporary register 21 and the calculated combined length unit / degree = 33 are included, and the dagger implement 23 is rotated out. That is the length of the first synthesized sound box: the length of the two synthesized sound boxes is different, and the length of the next two synthesized sound boxes is different. The register 2 1 will be equal to-= 3 厶.) The receiver (9 bits) is loaded into the subtraction. In the device 2 and 2, before the subtraction: the length of the sound box is deducted and the preset value of the sound box length is constant by the adder 23 = two =, degree (L), = 180), that is, + the mountain, ^ (Pre-revised value ^ number Shi Xiang) The length of the next sound frame U-M-Uum is determined by the parameters carried by the loading parameter unit, then U— = 180. The smoothing unit is used by the receiver.

Page 12 535140 V. Description of the invention (ίο) 3 0 For smoothing, please refer to "Figure 4". Among them, ', the smoothing processing unit 3 0 π includes a pitch parameter smoothing unit 3 and a calculation ratio unit 3 2, a calculation and synthesis length unit 33, an amplitude parameter smoothing unit 34, a spectral parameter smoothing unit 35, and a register. 3 6. After receiving the parameter data of the sound box twice, the smoothing processing unit 30, that is, the phoneme parameters (RC j (丨 0), rMS j, Pi tch j) of the sound box and the phoneme parameter of the previous phoneme (RC0 (10), RMS0, and Pitch0), that is, smoothing processing is started, and smoothing processing is performed every other base cycle length (pit ch). First, the proportional parameter is calculated from the calculating proportional unit 32, that is, Prop-L / M. Then the 'base period parameter (pj; tchj, PitchO) is processed by the pitch parameter smoothing unit 31 to obtain the processed base period, that is: Pitch = Pitch〇 * (l-Pr〇p) + Pitchj * Prop, the calculated base period (Pitch) can be sent to the temporary register 36 for temporary storage. The amplitude parameter (RMSj ·, RMS0) is smoothed by the amplitude parameter smoothing unit 34 to obtain the smoothed amplitude parameter, that is, RMS = RMS0 * (1-Pr〇p) + RMSj * Prop, which is also sent. It is temporarily stored in the temporary register 36. The smoothing processing of the spectral parameters (RC j (1 0), RC 0 (1 0)) is performed by the spectral parameter smoothing processing unit 3 5 to obtain the smoothed spectral parameters, that is, RC (i) = RC0 ( i) * (l-Prop) + RCj (i) * Prop, i = 〇, l, ..., 9, Similarly, the processed spectrum parameters are stored in the register 36. The pitch parameters, amplitude parameters, and spectral parameters stored in the register 36 are sent to the next part, that is, after the synthesis unit 40 is synthesized, the synthesis unit 40 synthesizes the parameters of the base cycle and sent out " One week

Page 13 535140 V. Description of the invention (11) " (Next_P itch) signal, this signal can control the output of the register 36. When the register receives this signal, it will load the smoothed signal of the next base cycle. Processed voice parameters. The π smoothing processing unit 3 0 ”calculates the synthesized length unit 33. After receiving the“ next base period ”(1 ^ 11: _? 丨 悦} 1) signal, it calculates the sound frame length of this synthesis, and obtains L = L + Pitch, if L > M then L = 0; Otherwise, send the signal of next frame n (N ext _ Frame) to the loading parameter unit 20, load the next frame parameter, and let L = 0. Among them, the initialization signal (initia 1) sent by the initialization unit 10 is sent to the calculation length unit 3 3, and L = 0 is used to initialize the unit. The next work is performed by the synthesis unit 40, please refer to "Figure 5", which contains the pulse sequence generator 41, the channel filter 4, the amplitude adjustment unit 43, and the memory 44. The pulse sequence generator 41 outputs a periodic pulse signal. This pulse signal is a waveform that simulates the vibration of the human vocal cord. It is stored in the memory contained in advance and its previous length is Pi tch. If Pi tch It is longer than the length of the pulse sequence stored in the memory. For example: the pulse sequence stored in the memory is {p [1], P [2], ..., p [2 5]}, then if Pitch > 25, output e (n) two {p [l], p [ 2],…, p [25], 0,…, 0} If Pitch < = 25, output e (n) = {p [l], p [2], ..., p [P it ch]} ° The channel filter 4 2 simulates the human oral cavity, nasal cavity, channel, etc. The resonance effect on the vocal fold vibration can be an all-pole filter (A 1 1 Po 1 e F i 11 er) or a point array wave filter (L a 11 ice F i 11 er), the input filter parameters are RC (i), i = 0, 1, 2, ... 9.

Page 14 535140 V. Description of the invention (12) After passing the pulse sequence through the channel filter 4 2 and then passing through the amplitude adjustment unit 43, the speech signal can be synthesized. The amplitude adjustment unit 43 calculates the amount of amplitude adjustment required by the RMS. After the speech synthesis, the amplitude adjustment unit 43 sends π the next base cycle n (Next_Pitch) to the smoothing processing unit 30. The memory 44 is responsible for temporarily storing the speech signals calculated by the channel filter 42 and the amplitude adjustment unit 43. Finally, the parameters processed by the smoothing unit 30 are synthesized into a basic periodic speech by the synthesizing unit 40 and sent from the memory 44 to the speech output unit 50 to output the speech. The speech output unit 50 has at least one memory buffer (Memory Buffer), so that each synthesized speech period is stored therein. Although the present invention is disclosed above with the foregoing preferred embodiments, it is not intended to limit the present invention. Any person skilled in the relevant arts can make some changes and retouch without departing from the spirit and scope of the present invention. The scope of patent protection of an invention shall be determined by the scope of patent application attached to this specification.

Page 535140 Brief Description of the Drawings Figure 1 is a system architecture diagram of the speech phoneme decoder of the present invention; Figure 2 is a specific embodiment of the speech phoneme decoder of the present invention; and Figure 3 is a speech phoneme decoding of the present invention FIG. 4 is a structural diagram of a smoothing processing unit in a specific embodiment of a speech phoneme decoder according to the present invention; and FIG. 5 is a speech phoneme decoder of the present invention. Architecture diagram of the synthesis unit in the specific embodiment. [Illustration of symbol and number] 10 Initialization unit 20 Load parameter unit 21 Register 22 Subtracter 23 Adder 24 Data load unit 241 Parameter calcite horse 25th --- Register 26th --- Temporary Register 30 Smoothing processing unit 31 Yinan parameter smoothing processing unit 32 Calculation ratio unit 33 Calculating synthetic length tic? Early element 34 Amplitude parameter smoothing processing unit 35 Spectrum parameter smoothing processing unit

Page 16 535140

Brief description of the drawing 36 40 41 42 43 44 50 8u 8 s 16u 16s initial LM Next_F rame Next_P i tch Pitch P itchj PitchO Prop RC RC j RC〇RMS register synthesizer unit pulse sequence generator channel filter amplitude adjustment unit Memory speech output unit unsigned 8 bits (signed 8 bits) signed fixed 8 bits (unsigned 16 bits signed 16 bits (signed 16 bits) Initialization signal has been synthesized length Synthesized sound frame length Next sound frame (Next_Frame) Next base cycle (Next_Pitch) Base cycle This processed base week The last processed base cycle ratio {Column (Proportion) Spectrum parameters this time Processed Spectrum Parameters Last Processed Spectrum Parameter Amplitude Parameter (root mean square value)

Page 17 535140

Page 18

Claims (1)

535140 6. Scope of patent application 1. A speech phoneme decoder, which uses an amplitude parameter (RMS), a base period parameter (Pi tch) and a linear predictive coding method (Linear P redictive Coding, LPC). ) To decode the speech data encoded by the spectral parameters (RC's), the encoded speech data is stored in a speech database, and the speech phoneme decoder includes: an initialization unit for generating an initialization signal; The input parameter unit is connected to the initialization unit, and is used to receive the initialization signal, and load a voice frame's voice data from the voice database with a frame (Frame) as a unit; a smoothing processing unit To receive the voice data of the voice frame, and to use one base period (Pi tch) of the voice frame as the length, use the internal difference method to separately process the amplitude in the voice data of the voice frame Parameters, the base period parameter and the spectrum parameter, send a signal of a sound frame to the loading parameter unit to load the voice data of the next sound frame; a synthesis unit, using Receiving the speech data of the base cycle processed by the smoothing processing unit and synthesizing the speech data into a speech signal; after the synthesis unit has processed the speech data of the base week, sending out a signal of the base cycle to the smoothing processing unit to process the next signal A base period of speech data; and a speech output unit for receiving a speech signal transmitted by the synthesis unit to output speech. 2. The speech phoneme decoder according to item 1 of the scope of the patent application, wherein the loading parameter unit includes a parameter decoder which decodes and decodes the code based on the base cycle parameter, the amplitude parameter, and the spectral parameter encoding order. Output to the smoothing unit in parallel. 535140 VI. Application for patent scope 3. The speech phoneme decoder described in item 1 of the scope of patent application, wherein the loading parameter unit is after loading the voice data of the voice frame, temporarily storing the voice of the current voice frame Data, and after receiving the signal of the next sound box transmitted by the smoothing processing unit, load the speech data of the next sound box, and combine the speech data of the current sound box with that of the next sound box The speech data is sent to the smoothing processing unit. 4. The speech phoneme decoder according to item 1 of the scope of patent application, the smoothing processing unit processes the speech data of the current sound frame and the speech data of the next sound frame by the internal difference method and outputs it to the synthesis unit. 5. The speech phoneme decoder according to item 1 of the scope of patent application, the smoothing processing unit includes: a calculation ratio unit for calculating the length of the synthesized sound frame of the current sound frame and the synthesized sound of the sound frame Proportion of frame length; a pitch parameter smoothing processing unit for receiving a base period parameter of the current sound frame and a base period parameter of the next sound frame, and calculating a composite base period parameter by an internal difference method; an amplitude A parameter smoothing processing unit is configured to receive the amplitude parameter of the current sound frame and the amplitude parameter of the next sound frame, and calculate a composite amplitude parameter by the internal difference method; a spectrum parameter smoothing processing unit is used to receive the current The spectrum parameter of the sound frame and the spectrum parameter of the next sound frame, and a composite spectrum parameter is calculated by the internal difference method; a calculation length unit is used to calculate the length of the synthesized sound frame of the current sound frame and input the result To the calculation scale unit and output this Page 20 535140 VI. The signal of the next sound frame in the scope of patent application to the loading parameter unit; and a temporary register to store the synthetic base cycle parameter, the synthetic amplitude parameter and the synthetic spectrum parameter and output to the Synthesis unit. 6. The speech phoneme decoder according to item 1 of the scope of patent application, the synthesis unit comprising: a pulse sequence generator for outputting the base period parameter as an excitation signal (Excitation Signal); A channel filter for receiving the excitation signal and processing it into a synthetic speech signal according to the spectral parameter as a filtering parameter of the channel filter; and an amplitude adjusting unit for multiplying the synthetic speech signal by The amplitude parameter is used to output a restored voice and output to the voice output unit. 7. The speech phoneme decoder according to item 1 or 6 of the scope of patent application, the synthesis unit further includes a memory for temporarily storing the synthesized speech signal and the restored speech, and outputting the restored speech to the speech output unit. Page 21