GB2322778A

GB2322778A - Noise output for a decoded speech signal

Info

Publication number: GB2322778A
Application number: GB9704316A
Authority: GB
Inventors: Dominic Sai Fan Chan
Original assignee: Motorola Ltd
Current assignee: Motorola Solutions UK Ltd
Priority date: 1997-03-01
Filing date: 1997-03-01
Publication date: 1998-09-02
Anticipated expiration: 2017-03-01
Also published as: DE19804557A1; GB2322778B; FR2760285A1; GB9704316D0

Abstract

A speech decoder (201) is arranged and constructed to receive a plurality of speech parameters and decode the plurality of speech parameters into at least one fragment of decoded speech. A noise generator (205) is arranged and constructed to output a noise signal. Determining means (203) provide a decision as to whether the plurality of speech parameters represents unvoiced speech. A switch (209), operating together with the speech decoder, the noise generator, and the means for determining, is arranged and constructed to output the noise signal when the plurality of speech parameters represents unvoiced speech. A noise source (2051) e.g. a pseudo-random sequence generator applies a Gaussian noise signal to an LPC filter (2053) which spectrally shapes the noise signal with an estimate of the original speech spectrum envelope as provided by the LPC parameters provided by speech decoder (201). A filter (2055) adjusts the output energy of the noise signal to match the original decoded speech.

Description

2322778 1 NOISE OUTPUT FOR A DECODED SPEECH SIGNAL

Field of the Invention

This invention relates generally to speech coding, including but 10 not limited to decoding of received speech parameters.

Background of the Invention

Communication resources such as radio frequency channels are, at least at the present time, limited in quantity. Notwithstanding this limitation, communication needs continue to rapidly increase. Dispatch, selective call, and cellular communications, to name a few, are all being utilized by an increasing number of users. Without appropriate technological advances, many users will face either impaired service or possibly a complete lack of available service.

Recent technological advance intended to increase the efficiency of data throughput, and hence decrease system capacity needs to thereby allow more communications to be supported by the available limited resources, is speech coding. Code Excited Linear Prediction (CELP) speech coders, Algebraic Code Excited Linear Prediction (ACELP) speech coders, and Vector Sum Excited Linear Prediction (VSELP) speech coders (the latter being a class of CELP coders) have been proposed that exhibit good performance at relatively low data rates. Rather than transmitting the original voice information itself, or a digitized version thereof, such speech coders utilize linear prediction techniques to allow a coded representation of the voice information to be transmitted instead. Utilizing the coded representation upon receipt, the voice message can then be reconstructed. For a general description of one version of a CELP approach, see United States Patent No. 4,933,957 issued to

2 Bottau et al., which describes a low bit rate voice coding method and system.

CELP type speech coders derive an excitation signal by summing a long term prediction vector with one or more codebook vectors, with each vector being scaled by an appropriate gain prior to summing. A linear predictive filter receives the resultant excitation vector and introduces spectral shaping to produce a resultant synthetic speech. Properly configured, the synthetic speech provided by such a speech coder will realistically mimic the original voice message.

A method of decoding speech includes taking speech parameters and generating an LPC (Linear Predictive Coding) filter 107 and an excitation to the filter from the speech parameters. The excitation can be categorized as voiced excitation or unvoiced excitation. Voiced excitation 101 refers to an excitation signal with repetitive pattern. This signal represents the periodic pulses of air which excite the vocal tract. The period of repetition represents the vibration frequency of the vocal folds. Unvoiced excitation 103 refers to an excitation signal with random sequence. This signal represents the turbulence generated at a constriction at some point of the vocal tract by a high velocity air flow through the vocal folds that do not vibrate but stay open. The excitation signal, a sum 105 of the voiced excitation signal and the unvoiced excitation signal, is input into an LPC filter 107 that generally represents the estimated original speech spectrum envelope. The LPC filter 107 outputs decoded speech.

Because speech parameters are compressed representations of speech, they may not adequately be decoded by a receiver into the voice or other sound(s) that were presents at a microphone of the transmitting device. In some situations, audible anomalies, such as clicks or other undesirable sounds, may be reproduced at the speaker of the receiver because the transmitted signal may not have enough bandwidth to appropriately transmit data representing the sounds at the microphone.

3 Accordingly, there is a need for a method of handling speech coding such that unwanted sounds or other audible anomalies are not present in the decoded speech without requiring more bandwidth.

Brief Description of the Drawings

FIG. 1 is a block diagram of a basic speech decoder.

FIG. 2 is a block diagram of a speech decoder disposed in a receiver in accordance with the invention.

FIG. 3 is a flowchart showing a method of outputting a noise signal as part of a decoded speech signal in accordance with the invention.

FIG. 4 is a flowchart showing a method of generating a noise signal in accordance with the invention.

is FIG. 5 is a flowchart showing a method of outputting a noise signal in place of unvoiced parts of a decoded speech signal in accordance with the invention.

Description of a Preferred Embodiment

The following describes an apparatus for and method of generating a noise signal to replace the unvoiced output for a decoded speech signal. The method comprises the steps of receiving a signal comprised of a voiced part and an unvoiced part; decoding the signal into a decoded voice part and a decoded unvoiced part; and outputting a noise signal instead of the decoded unvoiced part. A method of providing a decoded signal fragment comprises the steps of receiving a plurality of parameters representing a fragment of a signal, which parameters represent voiced or unvoiced speech; determining if the fragment of the signal is unvoiced; when the fragment of the signal is unvoiced, generating a noise signal as the decoded signal fragment.

A block diagram of a speech decoder that provides enhanced decoder speech within a receiver 200 is shown in FIG. 2. The preferred embodiment of the block diagram of FIG. 2 resides in a 4 DSP56300, available from Motorola, Inc. The receiver 200 is a radio frequency receiver in the preferred embodiment. Speech parameters, such as ACELP (algebraic code excited linear prediction) parameters as used in the preferred embodiment, are input into a speech decoder 201, which decoder decodes the ACELP parameters into decoded speech. In the preferred embodiment in the speech decoder, a frame of ACELP parameters is used to generate LPC parameters that specify the transfer function of an LPC filter and an excitation to the filter. In particular, the Line Spectrum Frequency codebook index in the speech parameters is transformed into LPC parameters that specify the LPC filter transfer function. The transfer function of the LPC filter is interpolated in the Line Spectrum Frequency (LSF) domain within 4 subframes. In the preferred embodiment, each 30 ms interval of speech is represented by a set of ACELP parameters and is referred to as a frame. Each frame comprises 240 samples of speech taken at an 8 kHz sampling frequency. Each 30 ms frame consists of 4 sub-frames, each of 7.5 ms duration comprising 60 samples. The excitation signal for each sub-frame is comprised of a codeword from the adaptive codebook and a codeword from the innovative codebook. The codewords are gain shaped by the gains derived from a codeword of the gain codebook and then combined to form the excitation signal. The excitation is input on a sub-frame by sub-frame basis to the LPC filter that outputs the decoded speech.

The speech parameters are also input to an unvoiced decision block 203, which determines whether or not the particular part of the signal currently being processed by the speech decoder represents a voiced or unvoiced part of the speech. Numerous methods of deciding whether a fragment of speech is unvoiced exist in the art. For example, methods based on threshold analysis on a few basic parameters are disclosed in the articles "A PitchSynchronous Digital Feature Extraction System for Phonemic Recognition of Speech" by W. J. Hess from IEEE Transactions on Acoustics, Speech, and Signal Processing, ASSP-24, No. 1, pp. 14-25 (1976) and "Automatic Discrimination of Noisy and Quasi Periodic Speech Sounds by the Phase Plane Method" by B. M. Lobanov from Sov.

Phys. Acoust. 16:353-356 (1970). Other methods are based on pattern recognition, such as those disclosed in the articles "A Pattern Recognition Approach to Voiced-Unvoiced-Silence Classification with Applications to Speech Recognition" by B. S. Atal, and L. R. Rabiner from IEEE Transactions on Acoustics, Speech, and Signal Processing, ASSP-24, No. 3, pp. 201-212 (1976) and "A Procedure for Using Pattern Classification Techniques to Obtain a Voiced/Unvoiced Classifier" by L. J. Siegel from IEEE Transactions on Acoustics, Speech, and Signal Processing, ASSP-27 No. 1, pp. 83-89 (1979). The above articles are hereby incorporated by reference. Thus, a voiced/unvoiced decision may be made by analyzing decoded speech.

In the preferred embodiment, the first reflection coefficient derived from the ACELP LPC parameters is used to make the voiced/unvoiced decision. Reflection coefficients may be obtained from LPC parameters using a backward recursion technique, such as is described on p. 444 of the book "Digital Processing of Speech Signals" by L. R. Rabiner and R. W. Schafer (1978). A recursion formula is:

ki ai(' + l_k i 2 for 1:5 j:5 i - 1 where ki are the reflection coefficients and a, are the LPC coefficients. The recursion starts from i=p,p-1,.-J. Note: p is the order of the original LPC polynomial. When a sub-frame has a first reflection coefficient value below a threshold, the sub-frame is considered unvoiced. The threshold is empirically determined by observing the values of the first reflection coefficient at the boundaries of unvoiced segments.

The output of the unvoiced decision block 203 provides a control input to a switch 209 that outputs either voiced (when the unvoiced decision block output is negative) or unvoiced data (when the unvoiced decision block output is positive) as enhanced decoded speech. One input to the switch 209 is the voice decoder 201 output, 6 and the other input to the switch is an output from a noise generator 205. When the unvoiced decision block 203 determines that the speech parameters represent unvoiced data, the switch 209 outputs the noise generator 205 output in place of decoded speech.

Numerous methods of generating noise exist in the art. In the preferred embodiment of the present invention, a noise source 2051 outputs a Gaussian noise signal to an LPC filter 2053. The preferred embodiment utilises feedback shift registers to produce a pseudo-random sequence. In particular, a 16- or 32-bit linear feedback shift register (LFSR) may be used. A number of outputs from the LFSR are then averaged to form the Gaussian noise signal that is input to the LPC filter 2053. In the preferred embodiment, 32 pseudo-random sequences are averaged to form the Gaussian noise signal. Other methods include generating uniform distributed random numbers, such as are provided by the UNIX rando function, or generating uniform distributed floating point numbers, such as are provided by the UNIX drand48() function, and averaging a number of the outputs of these functions to yield a Gaussian noise signal. The LPC filter 2053 receives the LPC parameters from the speech decoder and spectrally shapes the output signal from the noise source 2051 with an estimate of the original speech spectrum envelope as provided by the LPC parameters.

In the preferred embodiment, the output of the LPC filter is input to an energy matching filter 2055 that adjusts the output energy of the noise signal, as output from the LPC filter, to match the original decoded speech for the sub-frame being replaced with noise, and also smoothes out the variations of frame energy between successive unvoiced speech sub-frames. In other words, the energy signature.from the decoded speech is used as the energy signature for the noise signal, thus the energy signature for the noise signal is adjusted to substantially match the energy signature of the original decodes speech that the noise signal will replace in the speech coder output. The energy signatures need not exactly match, as the shape, frequency, and/or amplitude of the energy signature may vary around 20% while still obtaining enhanced decoded speech. An alternative 7 method of energy matching is a "linear averaging" of the noise signal, where energies of consecutive unvoiced segments are "linearly averaged" using a small order moving average filter. The energy signature of the noise between the beginning and end of the unvoiced segment is a linear function. Other non-linear functions may also be used to provide the energy signature of the noise between the end points. Linear averaging and other alternative methods of providing an energy signature for the output noise signal may not provide as all the desirable effects of the method used in the preferred embodiment and may yield other undesirable effects such as cancelling out plosives.

A flowchart showing a method of replacing unvoiced parts of speech signal with noise is shown at the flow chart of FIG. 3. A fragment of a signal is a part of a signal. In the preferred embodiment, a fragment is referred to as a sub-frame or a speech sub-frame. At step 301, a signal is received, which in the preferred embodiment is received in the form of ACELP parameters. A frame of ACELP parameters consists of an index to the Line Spectrum Frequency codebook that specifies the LPC filter and for each sub frame, a pitch delay index, an index to the innovative codebook specifying the position of the 4 pulses, a global sign flag of the pulses, a shift flag to indicate if the pulses should be shifted to the right by one sample, and an index to the gain codebook specifying the gains of the adaptive codeword and the innovative codeword. The part of the incoming signal that is presently being processed is analysed to determine if the part is voiced or unvoiced at step 303.

If the part is unvoiced at step 303, the process continues with step 305, where a noise signal is generated and output at step 307, and the process continues with step 311. If at step 303, the part is voiced, the process continues with step 309, where decoded voice is output as is conventionally done in the art. If there are more parts of a received signal to be processed at step 311, the process continues with step 303, otherwise the process ends.

A method of generating a noise signal, as may be utilised in noise generator 205 of FIG. 2, is shown in the flowchart of FIG. 4.

8 is Step 401, pseudo-random noise is generated, as described above, in the preferred embodiment. At step 403, the generated noise is applied to an LPC filter, which spectrally shapes the noise with the estimated original speech spectrum envelope, as described above. At step 405, the energy level of the unvoiced noise signal is matched to the decoded speech energy levels, such as is described with respect to the energy matching filter 2055 of FIG. 2. Many other methods of generating a noise signal exist, which methods may be successfully used in practicing the present invention. For example, any random number sequence with repetitive period large enough, e.g., with a frequency less than or equal to 0.1 Hz, will be adequate. Alternatively, the generation of noise signals may comprise storing and retrieving a number of digital noise signals, although such a method is costly in terms of memory.

An alternative method of replacing unvoiced speech with a noise signal in a decoded speech signal is shown in the flowchart of FIG. 5. At step 501, a signal is received and decoded into voiced and unvoiced parts at step 503. Although the preferred embodiment utilises ACELP speech coding and decoding, successful practice of the present invention is independent of the type of speech coder/decoder, also referred to as a speech codec. Thus, the present invention may be applied to VSELP, CELP, and other speech codecs. At step 505, a noise signal is generated and output in place of the decoded unvoiced part of the signal. In addition, the decoded voice parts are output normally.

Unvoiced speech is replaced with noise on a sub-frame by sub-frame basis in the preferred embodiment. In addition, all unvoiced fragments or subframes of a speech signal need not be replaced with a noise signal. For example, sub-frames with excitations having high-frequency energy above a certain threshold may not be replaced with a noise signal.

The present invention improves the speech quality of digitally encoded and decoded speech (ACELP speech in the preferred embodiment) without changing the speech encoder structure and without requiring additional transmission bandwidth. The improved 9 codec is therefore backward compatible with the original codec. The present invention also improves the synthesis of unvoiced speech at the decoder. The method used is based on a speech production model in which unvoiced speech is generated by a noise source input to an LPC digital filter. When an unvoiced segment of speech is detected either from the original decoded speech or from the synthesis parameters, the original decoded speech segment is replaced by another signal of the same energy. This signal is the output of the original LPC filter driven by a pseudo- random noise source. As a result, unwanted sounds such as clicks or other audible anomalies are not reproduced at the speaker of a receiver.

Claims

What is claimed is:

Claims 1 A method comprising:

receiving a signal comprised of a voiced part and an unvoiced part; decoding the signal into a decoded voice part and a decoded unvoiced part; and outputting a noise signal instead of the decoded unvoiced part
2. The method of claim 1, further comprising the step of, when the fragment of the signal is voiced, decoding the fragment of the signal into the decoded signal fragment.
3. The method of claim 1, wherein the signal comprises a plurality of parameters that are algebraic code excited linear prediction parameters.
4. The method of claim 1, further comprising the step of transforming a plurality of linear predictive coefficients (LPCs) into at least one reflection coefficient, wherein the LPCs are at least some of the plurality of parameters.
5. The method of claim 4, wherein at least one reflection coefficient is used to determine if the signal is unvoiced.
6. The method of claim 1, wherein a decoded part of the signal is used to determine if the signal is unvoiced.
7. The method of claim 1, further comprising the step of outputting filtered noise using a plurality of linear predictive coefficients.
8. The method of claim 1, further comprising the step of generating a noise signal comprising a pseudo-random noise signal.
9. The method of claim 8, wherein the pseudo-random noise signal is input into a filter.
10. The method of claim 1, wherein the decoded unvoiced part has a first energy signature and the noise signal has a second energy signature that is substantially similar to the first energy signature.
11. A method of providing a decoded signal fragment comprising: receiving a plurality of parameters representing a fragment of a signal, which parameters represent voiced or unvoiced speech; determining if the fragment of the signal is unvoiced; and when the fragment of the signal is unvoiced, generating a noise signal as the decoded signal fragment.
12. The method of claim 11, further comprising the step of, when the fragment of the signal is voiced, decoding the fragment of the signal into the decoded signal fragment.
13. The method of claim 11, wherein the plurality of parameters are algebraic code excited linear prediction parameters.
14. The method of claim 11, further comprising the step of transforming a plurality of linear predictive coefficients (LPCs) into at least one reflection coefficient, wherein the LPCs are at least some of the plurality of parameters.
15. The method of claim 14, wherein the determining step comprises using the at least one reflection coefficient to determine if the signal is unvoiced.
16. The method of claim 11, further comprising the step of decoding the fragment of the signal to determine if the signal is unvoiced.

12
17. The method of claim 11, further comprising the step of outputting filtered noise using a plurality of linear predictive coefficients.
18. The method of claim 11, wherein the step of generating the noise signal comprises producing a pseudo-random noise signal.
19. The method of claim 18, wherein the pseudo-random noise signal is input into a filter.
20. The method of claim 1, wherein the fragment of the signal has a first energy signature and the noise signal has second energy signature that is substantially similar to the first energy signature.

13
21. An apparatus comprising: a speech decoder, arranged and constructed to receive a plurality of speech parameters and decode the plurality of speech parameters into at least one fragment of decoded speech; 5 a noise generator, arranged and constructed to output a noise signal; means for determining whether the plurality of speech parameters represents unvoiced speech; and a switch, operating together with the speech decoder, the noise generator, and the means for determining, arranged and constructed to output the noise signal when the plurality of speech parameters represents unvoiced speech.
22. The apparatus of claim 21, wherein the noise generator 15 comprises: a noise source, arranged and constructed to provide a pseudorandom sequence; a filter, operating together with the noise source, arranged and constructed to process the pseudo-random sequence into a filtered 20 noise sequence; and means, operating together with the filter, for energy matching an energy level of the filtered noise sequence with an energy level of the at least one fragment of decoded speech, thereby yielding the noise signal.
23. The apparatus of claim 21, wherein the apparatus is disposed in a radio frequency receiver.

14
24. A speech decoder substantially as hereinbefore described in the accompanying drawings.
25. A method of providing a decoded signal fragment substantially as hereinbefore described in the accompanying drawings.
26. A method of decoding speech signals substantially as 10 hereinbefore described in the accompanying drawings.
27. A receiver substantially as hereinbefore described in the accompanying drawings.