Signal quantiser for speech coding

0001This invention relates to a quantizer and more particularly to an improved signal quantizer suitable for use in speech coding.

BACKGROUND OF THE INVENTION

0002Many signals exhibit a slowly-varying, predictable behavior over time, so that changes in the amplitude of the signal are typically small for consecutive time samples. For these signals, quantizers can be developed which are more efficient than simple scalar quantization. For example, a quantization method has previously been proposed in an article by A. McCree and TB Barnwell III entitled "A Mixed Excitation LPC Vocoder Model for Low Bit Rate Speech Coding," IEEE Transactions on Speech Processing, July 1995, pp. 242-250, which reduces the bit rate for quantization of the gain term in Mixed Excitation Linear Predictive (MELP) speech coder.

0003As described in the above cited article (incorporated herein by reference), the gain term in the MELP coder typically changes slowly with time, except for occasional large excursions at speech transition such as the beginning of a vowel. Therefore, the gain can be quantized more efficiently when grouped into pairs than when each gain is individually quantized. For each pair of consecutive gain terms, the second term is encoded in the traditional way, with a 5-bit (32 level) uniform scalar quantizer covering the entire dynamic range of the gain signal. However, the first term is encoded using only 3 bits (8 levels) and a more limited dynamic range based on the already transmitted values for the second term and the previous value of gain. This method reduces the bit rate of the gain quantization in the MELP coder with no perceivable drop in the quality of the speech signal. This operation is illustrated in Fig. 1 where level 101 represents the level of the first term in the past frame and 102 represents the second term in the past frame, 103 represents the first term in the present frame and 104 represents the level in the second term in the present frame. The second term 102 with 32 levels covers the entire range 10 dB (decibel) to 77 dB, for example. The first term with 3 bits is from 6 dB above the maximum to 6 dB below the minimum of the levels of the neighboring second terms. The actual step size is 1/8th of the range. The step size is dependent on the levels of each frame.

0004Unfortunately, this quantizer does not perform well in the presence of noisy digital channels. When bit errors are introduced, the decoded value for the second gain term can contain very large errors due to the wide dynamic range of this signal, resulting in annoying "pops" in the output speech.

0005In accordance with one embodiment of the present invention, applicant presents an improved quantizer which results in better performance for noisy communication channels.

SUMMARY OF THE INVENTION

0006A method for quantizing a signal and a quantizer is provided by taking advantage of the expected input signal characteristics. In accordance with one embodiment of the present invention an improved quantizer is provided wherein the value of a first term representing a first time period is provided by first encoder and a value of a second term during a second adjacent time period is provided by a second encoder. The quantizer includes an encoder means responsive to a steady state condition to generate a special code. The quantizer includes a decoder with means responsive to the special code to provide an average of decoded second terms for a first term.

0007These and other features of the invention that will be apparent to those skilled in the art from the following detailed description of the invention, taken together with the accompanying drawings.

DESCRIPTION OF THE DRAWINGS

0008In the drawing: <ul id="ul0001" list-style="none"><li>Fig. 1 illustrates frames and first and second terms in frames;</li><li>Fig. 2 is a block diagram of a communications system;</li><li>Fig. 3 is a block diagram of an analyzer in the communications system of Fig. 2;</li><li>Fig. 4. is a block diagram of the improved quantizer according to one embodiment of the present invention;</li><li>Fig. 5 is a flow chart of the processor operation of Fig. 4;</li><li>Fig. 6 illustrates a synthesizer in accordance with one embodiment of the present invention;</li><li>Fig. 7 is a functional diagram of the decoder in the synthesizer of Fig. 6; and</li><li>Fig. 8 is a flow diagram for the operation of the decoder of Fig. 7.</li></ul>

DETAILED DESCRIPTION OF ONE EMBODIMENT OF THE PRESENT INVENTION

0009Human speech consists of a stream of acoustic signals with frequencies ranging up to roughly 20 KHz; however, the band of about 100 Hz to 5 KHz contains the bulk of the acoustic energy. Telephone transmission of human speech originally consisted of conversion of the analog acoustic signal stream into an analog voltage signal stream (e.g., by using a microphone) for transmission and conversion back to an acoustic signal stream (e.g., use a loudspeaker). The electrical signals would be bandpass filtered to retain only the 300 Hz to 4 KHz frequency band to limit bandwidth and avoid low frequency problems. However, the advantages of digital electrical signal transmission has inspired a conversion to digital telephone transmission beginning in the 1960s. Digital telephone signals are typically derived from sampling analog signals at 8 KHz and nonlinearly quantizing the samples with 8 bit codes according to the µ-law (pulse code modulation, or PCM). A clocked digital-to-analog converter and compounding amplifier reconstruct an analog electric signal stream from the stream of 8-bit samples. Such signals require transmission rates of 64 Kbps (kilobits per second) and this exceeds the former analog signal transmission bandwidth.

0010The storage of speech information in analog format (for example, on magnetic tape in a telephone answering machine) can likewise be replaced with digital storage. However, the memory demands can become overwhelming: 10 minutes of 8-bit PCM sampled at 8 KHz would require about 5 MB (megabytes) of storage.

0011The demand for lower transmission rates and storage requirements has led to development of compression for speech signals. One approach to speech compression models the physiological generation of speech and thereby reduces the necessary information to be transmitted or stored. In particular, the linear speech production model presumes excitation of a variable filter (which roughly represents the vocal tract) by either a pulse train with pitch period P (for voiced sounds) or white noise (for unvoiced sounds) followed by amplification to adjust the loudness.

0012The application cited above of A. McCree entitled "Mixed Excitation Linear Prediction with Fractional Pitch" filed March 3, 1994, incorporated herein by reference , is one such low bit rate speech coder. Figure 7 of that application describes the overall system and is shown herein as Figure 2. The input speech is sampled by an analog to digital converter and the parameters are encoded in the analyzer 600 and sent via the storage and transmission channel to the synthesizer 500 . The decoded signals are then converted back to analog signals by the digital to analog converter for output to a speaker. Fig. 5 of the reference application illustrates the synthesizer. The synthesizer is also illustrated in applicant's article in IEEE Trans. on Speech and Audio Processing Vol. 3 ,N0.4, July 1995 in an article entitled "A mixed Excitation LPC Vocoder Model for Low Bit rate Speech Coding".

0013Fig. 3 herein (like Fig. 6 in the above cited application)illustrates the analyzer 600. The analyzer 600 receives the analog speech and converts that to digital speech using analog to digital converter 620. The digitized speech is applied to an LPC extractor 602, pitch period extractor 604, jitter extractor 606, voiced /unvoiced mixture control extractor 608, and gain extractor 610. An encoder(controller)612 assembles the block outputs and clocks them out as a sample stream. The five arrows into encoder 612 are from the five output blocks. The encoder for gain 610 uses the quantizer system 612a according to the present invention.

0014As discussed in the background the encoder provides two output levels each frame . As illustrated in Fig. 1 the level 101 is the first term in the past frame and level 102 is the second term in the past frame. The levels 103 and 104 represent the first and second terms of the present frame. In order to reduce the bit rate of gain quantization the first term is encoder using only three bits and the second term by using five bits. The second term with 32 levels covers the entire range of levels from 10 dB to 77 dB. The first term with only three bits has a range of from 6 dB above the maximum to 6 dB below the minimum of the levels of the neighboring second terms.

0015As discussed in the background this system does not perform well in the presence of noisy digital channels. When bit errors are introduced, the decoded value for the second gain term can contain very large errors due to the wide dynamic range of this signal. This causes annoying "pops".

0016Applicants' new quantization method and system herein avoids this problem by taking advantage of the expected steady-state behavior of the gain signal. Typically, the gain does not vary much over the two terms in the current frame and the previous gain term. For these cases, we introduce a special quantization code for the first gain term to represent a steady-state frame. When the decoder in the synthesizer 500 detects the special steady-state code, it simply uses an interpolated value for the first gain term based on the second gain term value and transmitted second and previous gain value. The frames are stored before being transmitted so the second term of the same frame is available for first term calculation. This method improves the performance of the quantizer during steady-state frames under bit errors by introducing redundancy to the transmitted values. For these steady-state frames, the decoder will still produce a steady-state output during bit errors as long as at least some of the information bits are uncorrupted. As long as either the steady-state code or the second gain term is correctly received, the decoder will not introduce significant excursions into the output gain signal. During speech frames which are not steady-state, bit errors will tend to make the decoder of this new quantizer produce a smoother output than was intended. However, steady-state frames occur much more frequently and are more perceptually important than transitional frames, so the performance improvement obtained for the steady-state frames far outweighs any slight quality degradation in bit error performance for transitional frames.

0017This quantizer has been implemented in a 2.4 kb/s MELP coder, and compared against the previous quantization method. In informal listening tests, the new method provides a clear improvement during bit error conditions, while having no negative effect on clean channel performance.

0018Referring to Fig. 4 there is illustrated the encoder portion of the quantizer system according to one embodiment of the present invention. The input speech is converted to digital at A/D converter 620(Fig. 3) and applied to gain extractor 610. The speech gain is estimated twice per frame using an RMS power level detector of the input speech. For voiced frames the estimation window is adjusted to be a multiple of the pitch period. The gain output from extractor 610 is applied to encoder 612a. At encoder 612a the converted speech gain over the first half of each frame (first term) is switched via switch 704 to log function detector 703 and over the second half of the frame to log function 705. The log function 703 and 705 may be provided by a log look-up table where for a given gain level a log gain is provided. The output from log function 703 is applied to a 3-bit uniform (equal size step) scalar encoder 707 to provide the first term and the output from log function 705 is applied to 5-bit uniform scalar encoder 709 to provide the second term. The second term is for the whole range from minimum (10 dB for example) to maximum (76 dB the example) in 32 steps as represented by the 5-bits. The encoder 707 is capable of a range of 7 levels with three bits. The first term with only three bits has a range of from 6 dB above the maximum to 6 dB below the minimum of the levels of the neighboring second terms. The eighth possible level is reserved for a special code. A processor 710 is coupled to the 3-bit encoder and 5-bit encoder . The processor is coupled to storage 711. The processor 707 is programmed to follow the procedure of the flow chart of Fig. 5.

0019The processor stores (steps 801) the terms (gain levels) of the frames in storage 711. The second term of the current frame is compared (step 803) to the second term of the previous frame and the system determines if these two terms have a gain level within 5 dB of the other. If so (yes at step 805), the processor compares (step 807) the value of the first term of the current frame (the intermediate term) to average gain levels of the second terms of the current and previous frames (the halfway point) and if it is within 3 dB of the average gain level (807) then the special steady state code is sent from the 3-bit decoder. The second term (past and current) are averaged at step 811 and sent to comparator 807. If the conditions are not met, the system behaves as before with only seven levels available for the first term since the 8th level is used for the special code.

0020Referring to Fig. 6 there is illustrated a synthesizer according to one embodiment of the present invention. For the example in the present invention this is the MELP system as shown in Fig. 5 of the above cited application incorporated herein by reference.

0021Fig. 6 illustrates in functional block form a first preferred embodiment speech synthesizer, generally denoted by reference numeral 500, as including periodic pulse train generator 502 controlled by a pitch period input, a pulse train amplifier 504 controlled by a gain input, pulse jitter generator 506 controlled by a jitter flag input, a pulse filter 508 controlled by five band voiced/unvoiced mixture inputs, white noise generator 512, noise amplifier 514 also controlled by the same gain input, noise filter 518 controlled by the same five band mixture inputs, adder 520 to combine the filtered pulse and noise excitations, linear prediction synthesis filter 530 controlled by 10 LSP inputs, adaptive spectral enhancement filter 532 which adds emphasis to the formants, and pulse dispersion filter 534. Filters 508 and 518 plus adder 520 form a mixer to combine the pulse and noise excitations.

0022The control signals (LPC coefficients, pitch period, gain, jitter flag, and pulse/noise mixture) derive from analysis of input speech. Fig. 3 illustrated in functional block form a first preferred embodiment speech analyzer, denoted by reference numeral 600, as including LPC extractor 602, pitch period extractor 604, jitter extractor 606, voiced/unvoiced mixture control extractor 608, gain extractor 610, and controller 612 for assembling the block outputs and clocking them out as a sample stream. Sampling analog-to-digital converter 620 could be included to take input analog speech and generate the digital samples at a sampling rate of 8 KHz.

0023The encoded speech may be received as a serial bit stream and decoded into the various control signals by controller and clock 536. The clock provides for synchronization of the components, and the clock signal may be extracted from the received input bit stream. For each encoded frame transmitted via updating of the control inputs, synthesizer 500 generates a frame of synthesized digital speech which can be converted to frames of analog speech by synchronous digital-to-analog converter 540. Hardware or software or mixed (firmware) may be used to implement synthesizer 500. For example, a digital signal processor such as a TMS320C30 from Texas Instruments can be programmed to perform both the analysis and synthesis of the preferred embodiment functions in essentially real time for a 2400 bit per second encoded speech bit stream. Alternatively, specialized hardware (e.g., ALUs for arithmetic and logic operations with filter coefficients held in ROMs, RAM for holding encoded parameters such as LPC coefficients and pitch, sequences for control, LPC and LSP conversion and back special circuits, a crystal oscillator for clocking, and so forth) which may hardwire some of the operations could be used. Also, a synthesizer alone may be used with stored encoded speech. The encoded speech is applied to control and clock decoder 536 and the five outputs are the gain control, pitch period, extractor, jitter flag, LPC coefficients and voiced/unvoiced mix. The decoding occurs in control and clock decoder 536a. For the gain function, the decoder comprises the subsystem of Fig. 7. The encoded input is switched by switch 901 to 3-bit decoder 903 and 5-bit decoder 905 every half frame to provide the first and second terms. The decoder 903 contains, for example, a look-up table that for a 3-bit input term provides a given gain level. The decoder 905 for the 5-bit input term provides a given gain level. An anti-log function of value is calculated at 906 and 908 and provided as gain to the amplifier 504 and 514 in Fig. 6. The anti-log can also be provided by a look-up table. A processor 909 and memory 910 are coupled to the decoder 903 and 905.

0024The processor 909 stores the current and previous second term gain and averages these gains. The processor looks for the special code and if it does receive the special code as shown in Fig. 8 it averages the value of the received previous and current second gain terms and compares this to the previous gain and if within 5 dB provides the average value to the gain in the synthesizer. If not and if the previous frame was not in error, then it is assumed there was a bit error and a channel error counter is incremented. The previous second frame gain value is repeated for both terms of the current frame. To ensure that the decoder correctly tracks the encoder, the processor does not implement the repeat mechanism if the previous frame was in error.

0025The pseudo code for the encoder and decoder follows: <ul id="ul0002" list-style="none" compact="compact"><li>ENCODER</li><li>Code second gain term using 5-bit quantizer</li><li>If (second gain within 5 dB of previous value AND    first gain within 3 dB of interpolated value)    transmit special interpolation code for first gain</li><li>Else    code first gain with 7 remaining quantizer levels</li><li>Endif</li><li>DECODER</li><li>Decode second gain term</li><li>Decode first gain term</li><li>If interpolation code received and average of the received and previous second term gain is within 5 dB from previous second term gain provide average gain    If (interpolation code received AND    second gain more than 5 dB from previous gain AND    previous frame was not in error)    assume channel error was introduced to steady-state frame increment channel error counter    repeat previous gain value of second term clear channel error counter</li><li>Else    clear channel error counter</li><li>Endif</li></ul>

0026Although the present invention and its advantages have been described in detail, it should be understood that various changes, substitutions and alterations can be made herein without departing from the spirit and scope of the invention as defined by the appended claims.