TW410323B - Method of high-speed adaptive differential pulse code modulation for voice coding - Google Patents

Abstract

The voice coding method of high-speed adaptive differential pulse code modulation uses comparators, continuous approach and two or three kinds of table lookup for instance quantification table, quantification threshold table, etc. This method reduces operation time, memory capacity and simplifies hardware. The memory that stores voice is decreased down to 24Kbits per second or less. It can reach the most effective compression for voice coding and playback without any distortion.

Description

410323

------ Case No. 8711 Description of the 5 'invention (1) 2 Ming is a fast-adaptive differential pulse method for speech coding, especially using a comparator and continuous approximation method, plus 3 lookup tables (Such as quantization table, quantization critical table, etc.), which can save computing time. Reduce the amount of memory used. The hardware is very simple. The memory used to store the sound is reduced to 24Kbits / SeC or lower, which can edit the voice signal. The code b achieves the most efficient compression without distortion during playback ^ ..... The traditional PCM (Pulse Code Modulation, Pulse Code Modulation) code is used to encode the sigmoid signal. The values are independent and unrelated. In this way, the entire amplitude coding of the sampling points requires more bits', so the digitized voice signal bandwidth greatly increases. However, in fact, most of the voice signal sources are sampled at the Nyquist sampling rate or higher. The values of each sampling point are closely related. Big redundancy. With this correlation of the sources, a PCM encoding of the difference between the sample point values can be transmitted. In this way, under the condition that the quantization noise is unchanged, the number of coded bits is reduced, and the signal bandwidth is greatly compressed. This PCM coding is called Differential Pulse Code Modulation (DPCM). The disadvantage of DPCM is that when the sound is very low, the noise is loud, and when the sound is loud, the sound quality is not good; in the DPCM scheme, a predictor is used to predict the next prediction value, and the predictor can be adaptable. The difference between two adjacent signals and the quantization step are adaptability changes. Another type of adaptive ADPCM (Adaptative Differential Pulse Code Modulation) is to adapt the quantization step to change. Conventional ADPCM sound generators such as the Republic of China Patent Publication No. 119243 (the Republic of China Patent Application No. 77102181, the invention name

Throw sleeve 116435

(Named: ADPCM sound generator, applicant: Industrial Technology Research Institute), which revealed that a sound generator ’includes ADPCM mathematical logic circuit, mute reproduction circuit, syllable repeat control circuit, and read-only memory. This kind of sound generator can reduce the memory required for sound storage to 3 2 K bits / sec by ADPCM processing technology, and then reduce the memory to 20-26Kbits / sec by mute elimination and regeneration technology. For some animals, Sound, due to the repeatability of its syllables, the memory can be reduced. However, the circuit of this patented technology is complicated compared with the present invention, and the memory used is still large. The main purpose of the present invention is to provide a quick adaptive differential pulse code modulation of voice coding of speech coding, using a comparator and a succeyive apprach method 'SAM, and then In addition, the use of 2 or 3 types of lookup tables (1001 (_ 叩 tab 1 e, such as quantization tables, quantization critical tables, etc.) can save computing time. 'Reduce the capacity of the memory used. The hardware is very simple and stored. The memory required for the sound is reduced to 24 Kbits / sec, which can achieve the most efficient compression when encoding the speech signal without distortion during playback. The second object of the present invention is to provide a rapid adaptation differential pulse encoding for speech exhaustion. For the modulation method, when the continuous approximation method is performed, it uses a single comparator to compare the quantized value with the previous quantized value of the continuous approximation method, and the hardware is very simple. Another object of the present invention is to provide a fast adaptation to speech coding.

Page 5: 410323-Case was 8711B435 _ year, month and month amended _____________ 5. Description of the invention (3) The method of sub-pulse coding modulation, the continuous approximation method used is a binary search method, which is fast and convenient to search . Another object of the present invention is to provide a fast adaptive differential pulse coding modulation method for speech coding. When coding, the quantization table used may be a linear or noniinear quantization table. The non-linear quantization table is used Quantization table of log function ° Brief description of the diagram: FIG. 1 is a time function diagram of the speech signal of the present invention. FIG. 2A and FIG. 2B are graphs of a quantization table with a function of 10§ in a specific example of the present invention. This graph is a quantization table function diagram of a log function of 8 × 8. FIG. 3 is a quantization table of a specific example of the present invention. FIG. 4 is a schematic diagram of a comparator circuit of the present invention. FIG. 5 is a flowchart of a fast-adaptive differential pulse code modulation method for speech coding according to the present invention. Fig. 6 is a flowchart of steps for encoding and storing a pair of sampling points in the flowchart of Fig. 5 according to the present invention. FIG. 7 is a schematic diagram of a quantization threshold table of a specific example based on the quantization table in FIG. 3 of the present invention. Brief description of drawing number: 0… .start 1… .initialization 2… .encode a sample point and store the encoded value

410323

a Amendment V. Description of the invention (5) 513 ... Comparison result output 52 ... D / A converter 61 7 ... Quantization table content value 717 ... Quantization critical table content 611, 612, 613, 614, 615, 616 711, 712 , 713, 714, 715, 716 value detailed description of the invention: Please refer to Figure 1, this sample is timed sampling, false 0,1 at t1, the difference between the two signals is 0, check the amount of Figure 2A and Figure 2B That is, QI is Qi, i is 2, 0.9 is the difference between the two signals, and the next quantization index table lookup table is Q5, i is 3, or code. This method is the code modulation method of the present invention. The speech signal of the clear-speech coded is set to -0.1 in PCM, and the amplitude is 2 at t2. The difference of 0.2 is calculated by using the continuous approximation method to assume the value of index 2 of Q1, and also Q1 and I. The value is encoded at 13; the amplitude is .8 at 13 o'clock, and the difference of 0.8 is calculated by using the continuous approximation method as the value of index 3 of Q5. Quickly adapting to the differential pulse mapping two A and two B is a graph of a quantization table using a log function in a specific example of the present invention. This graph is a quantization table function chart of a 8x8 log function. There are 8 1 quantization tables from Q0 to Q7. Figure 3 shows an example of a quantization table, which shows how to compile the data into a quantized value according to the index. Regarding the fast-adaptive differential pulse code modulation method of the speech coding of the present invention, please refer to FIG. 5, which mainly includes the following steps:

IB1 410323

Fifth, the description of Mengming (6)%) Initialization Λ., ,, and 'Q1 ndex = 0, DacVa 1 ue = 80Η; ^ T sample point encoding and store the encoding value 2, the encoding is using the quantization ° bound table and Successive approach method (In a specific example of the present invention, the continuous approach method is a binary search method; (C) ▲ 々Record e code 3? If yes, then return to step (b), if not Then end 4; (D) Then the main result of the quantization critical table (QTT [Q]) is mainly added, and the positive and negative (direction) bits (si gn bit) are stored as the compressed data. And the above step (B) The detailed flow of encoding a sampling point and storing the encoded value 2 'refer to FIG. 6' includes the following steps: Step 22: Initiation, da — D / A, 0 — EN, 1 0 02 (binary ) — SM, the DA value is stored in d / a, 0 is stored in EN, and 002 is stored in SM;

Step 23 uses a comparator (the one in FIG. 4) to compare VI with…; if D / A > νι ', proceed to step 241, if D / A < VI, proceed to step 251; Step 241: C ( DA-QTT [QI] [EN + SM])-> D / A, after subtracting QTT [QI] [EN + SM] from the DA value, then take the function value of the clipping function (C1 ippingfuncti ο η), Deposit d / A; Step 242: Compare VI with D / A 'If VI < D / A, proceed to step 243' EN = EN + SM, add EN plus SM and then store in EN; then proceed to step 244; If VI &D; D / A, proceed to step 244; Step 244: SM shift one bit to the right, and then proceed to step 245, that is, divide the value of SM by 2, and then proceed to step 245;

Μ

Step 245: Determine whether SM is equal to 0. If SM is equal to ο, it means that 'to the nearest quantized value' has been forced to proceed to step 246: otherwise proceed to step 241; ^ step 2 4 6 • This step is used to synthesize the restored data , Used as the basis for the next storage encoding, its action is to store EN and minus sign, D ^ = C (DA-QT [QI] [EN], QI = NT [QI] [EN], where 212 疋 represents the direction This step is to subtract the value of qT [QI] [EN] from the DA value and then take the function value of the limiting function (Cl ipping function) into DA, and store the value of NT [QI] [ΕΝ] into qi , And then proceed to step 26 to return; Step 251: C (DA + QTT [QI] [EN + SM])-D / A, that is, the DA value is added to QTT [QI] [EN + SM], and then the limit is taken The function value of the function (Clipping functi〇n) is stored in D / A; Step 252: VI and D / A are compared, if VI> D / A, proceed to step 253, EN = EN + SM, that is EN plus SM Re-enter EN; then proceed to step 254; if VI < D / A, proceed to step 254; step 254: SM shift one bit to the right, then proceed to step 255, that is, divide the value of SM by 2 and then proceed Go to step 255 Step 2 5 5: Determine whether SM is equal to 0, if SM is equal to 0, it means that the nearest quantized value has been forced, proceed to step 2 5 6; if SM is not equal to 0, proceed to step 251; step 256: Store EN and positive sign (that is, compressed data), DA = C (DA + QT [QI] [EN]) 'QI = NT [QI] [ΕΝ]; that is, DA value plus QT [QI] [ΕΝ] after' Then take the function value of the Clipping function into DA, and store the value of NT [QI] [ΕΝ] into QI.

Page 10 410323 Case No. 87116435 Month_Revision_ V. Description of the invention (8) Return to step 26; Step 26: Return; all the above operations can be performed in the microcontroller. The above variables are abbreviated:

Qindex: QI: used to represent the QI quantization table Quantization Table: QT []: Quantization Table Encoded Value: EN: Coded Value Dac Value: DA: Dac Value Shift Mask: SM (also known as SAR): displacement screen Vin: VI: Input voltage

Quantization Threshold Table: QTT []; Quantization Threshold Table; QTT [N] [0]; QTT [N] [i] = iNT (QT [N] [I'1] + QT [N] [^] / 2) , Bu 1-7; 1 = Bu 7; 1 ^^ () is the rounding function; ^

Next Quantization Index Table: NT []: Next Quantization Index Table

Cl ipping function: C []: Limiting function: Assumed to be 8. bit system, when Ving255, C [Vin] = 255; when VinSO, C [Vin] = 0. FIG. 4 shows a circuit diagram of the present invention comparing VI with D / A in steps (23), (242), and (252). In the figure, the input of D / A converter 52 is a function value obtained in step 241 or step 251, and Vin is a PCM input voltage signal sampled by PCM, which is connected to Vin input terminal 511 of comparator 51, and D / A converter 52 stores D / A value, and D / A voltage value of D / A converter 52 is input to comparison

— 87ITd4j5 Month _ Day Amendment _ V. Explanation of the Invention (9) Another D / A input terminal 512 of the device 51 compares Vin with D / A, and the comparison result is output to the micro-controller through the comparison result output terminal 5 1 3 Device. FIG. 7 is a schematic diagram of a quantization critical table used in the present invention. The quantization table (QT [QI] [i], i = 〇, 1,2,3.'7) * 611, 612, 613, 614, 615, 616, 617, 618 Content values 0.1, 0.2, 0.4, 0.6, 0.9, 1.3, 1.9, 2.8 are the quantization table indexes (in (jex) i, 2, 3, 4, 5, 6, 7, The quantized value of 8 and the content value of 711 (qtt [Qi] [i] 'i = 〇, 1,2, 3, 4 ...) equal to 611 content value (oi) plus 6 1.2 content value (0, 2) divided Take 2 (equal to 0.15) 'Similarly, the content value of 712 is equal to the content value of 612 (0.2) plus the content value of 613 (0.4) divided by 2 (equal to 0,3); the content value of 713 is equal to the content value of 613 ( 〇4) plus 614 content value (〇 · 6) divided by 2 (equal to 0.5); 714 content value is equal to 614 content value (〇.6) plus 615 content value (0, 9) divided by 2 (equal to 〇, 75) ..., using the quantized critical table (qTT [qi]

[I]) 'When performing the binary search method, compare the size with Vin; if Vi should be to the right of QTT [QI] [i], look to the right of QTT [q 〖] [丨], otherwise, if Vi should be to the left of QTT [QI] [i], then go to the left of QTT [QI] [i]. You can use the quantization critical table to quickly find and find QI and i. One

Claims (1)

Patent scope 1. A kind of fast-adaptive differential pulse coding for speech coding includes: 3 η Rinai A. Initiation; Β. Coding a sampling point and storing the coding value, the coding is using a quantization table and a quantization critical table And continuous approximation (success appr0ach methQd); C. Continue recording encoding? If it is, then return to step (β), if not, end it, and 'σ D. Then, the final result of the quantization critical table (qtτ [q]) is added, plus the sign (sign) bit is The compressed data is saved. 2. The method of fast adaptive differential pulse code modulation for speech coding according to item 1 of the scope of patent application, wherein the continuous approximation method of step (B) is a binary search method. 3. For example, the speech coding of item 2 of the scope of patent application, a method of fast adaptive differential pulse code modulation, wherein step (β) includes: Step 22: Initiation, DA-D / A, 0-EN, 1 0 02 — SM, store the DA value in D / A, 0 in EN, and then store 002 in SM; Step 23 Use a comparator (the one in Figure 4) to compare VI with D / A; if D / A > VI 'then proceed to step 241, if D / ASVI, proceed to step 251; Step 241: C (DA-QTT [QI] [EN + SM])-D / A, that is, according to the current quantification The table (QTT [QI]) tries to compare the finer section (SM) in the current range (EN), subtracts this difference from DA, and then takes the function value of the clipping function and stores it D / A; Step 242: VI and D / A are compared, if VI < D / A, proceed to step 243, EN = EN + SM, add EN plus SM and then store in EN; then proceed to step 244; if VI > D / A, proceed to step 244; Patent scope on page 13 1. A kind of fast-adaptive differential pulse coding for speech coding includes: 3 η Rina harness A. Initiation; Β. Encoding and storing the encoding value of a sampling point, the encoding is using a quantization table, Quantization critical table and continuous approximation method (success appr〇ach methQd); C. Continue recording encoding? If it is, then return to step (β), if not, end it, and 'σ D. Then, the final result of the quantization critical table (qtτ [q]) is added, plus the sign (sign) bit is The compressed data is saved. 2. The method of fast adaptive differential pulse code modulation for speech coding according to item 1 of the scope of patent application, wherein the continuous approximation method of step (B) is a binary search method. 3. For example, the speech coding of item 2 of the scope of patent application, a method of fast adaptive differential pulse code modulation, wherein step (β) includes: Step 22: Initiation, DA-D / A, 0-EN, 1 0 02 — SM, store the DA value in D / A, 0 in EN, and then store 002 in SM; Step 23 Use a comparator (the one in Figure 4) to compare VI with D / A; if D / A > VI 'then proceed to step 241, if D / ASVI, proceed to step 251; Step 241: C (DA-QTT [QI] [EN + SM])-D / A, that is, according to the current quantification The table (QTT [QI]) tries to compare the finer section (SM) in the current range (EN), subtracts this difference from DA, and then takes the function value of the clipping function and stores it D / A; Step 242: VI and D / A are compared, if VI < D / A, proceed to step 243, EN = EN + SM, add EN plus SM and then store in EN; then proceed to step 244; if VI > D / A, proceed to step 244; Page 13 41 (m VI. Step 244 of patent application scope: SM shifts one bit to the right (SM = SM > > 1), and then proceeds to step 245, that is, the value of SM is divided by 2, and then proceeds to step 245; step 245: Determine whether SM is equal to 0, if SM is equal to 0, the table is forced to the nearest quantized value, proceed to step 246; otherwise, proceed to step 24j; step 246: store £ ^ and negative sign, 〇 = (: (0 Enter -〇1 '[卩 1] [£ 1 ^], QI = NT [QI] [EN]' Subtract the DA value from QT [QI] [EN] 'and then take the function of the Clipping function The value is stored in μ, and the value of NT [QI] [EN] is stored in QI, and then the process returns to step 26; Step 251: C (DA + QTT [QI] [EN + SM]) ~ &D; A , That is, {^ value plus QTT [QI] [EN + SM], and then take the function value of the clipping function and store it in D / A; Step 252 11 is compared with 0 / ^, if ¥ 1 > 0 / Person, proceed to step 253, EN = EN + SM, that is, EN plus SM and then store it in EN; then proceed to step 254; if VI < D / A, proceed to step 254; step 254: SM right Shift by one bit (SM = SM > > 1), then proceed to step Step 2 5 5 'divide the value of SM by 2 and proceed to step 255; Step 2 5 5: Determine whether SM is equal to 0. If SM is equal to 0, the table is forced to the nearest quantized value and proceed to step 256; otherwise proceed to step 251; step 2 5 6: store EN and minus sign, DA = C (DA + QT [Qi] [EN]), QI = NT [QI]. [ΕΝ]; that is, DA value plus QT [Ql] [εν], then take the function value of the clipping function and store it in μ, and store the value of NT [QI] [EN] in QI, and then go to step 26 to return. Page 14