EP0333425A2 - Sprachcodierung - Google Patents

Sprachcodierung Download PDF

Info

Publication number
EP0333425A2
EP0333425A2 EP89302481A EP89302481A EP0333425A2 EP 0333425 A2 EP0333425 A2 EP 0333425A2 EP 89302481 A EP89302481 A EP 89302481A EP 89302481 A EP89302481 A EP 89302481A EP 0333425 A2 EP0333425 A2 EP 0333425A2
Authority
EP
European Patent Office
Prior art keywords
pitch
sequence
output
filter
lpc
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
EP89302481A
Other languages
English (en)
French (fr)
Other versions
EP0333425A3 (de
Inventor
Barry G. Dept. Of Electronic & Elect. Eng. Evans
Ahmet M. Dept. Of Electronic & Elect. Eng. Kondoz
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
University of Surrey
Original Assignee
University of Surrey
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by University of Surrey filed Critical University of Surrey
Publication of EP0333425A2 publication Critical patent/EP0333425A2/de
Publication of EP0333425A3 publication Critical patent/EP0333425A3/de
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L19/00—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis
    • G10L19/04—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis using predictive techniques
    • G10L19/06—Determination or coding of the spectral characteristics, e.g. of the short-term prediction coefficients
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L19/00—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis
    • G10L19/04—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis using predictive techniques

Definitions

  • the present invention relates to speech coders and, more particularly, to low bit rate speech coders.
  • the invention also relates to a method of coding speech for transmission along a telecommunications link.
  • Speech is a complex analogue waveform.
  • the information contained in the analogue signal must be reduced to information in digital form.
  • This technique is known as speech coding.
  • the analogue speech signal is sampled and a digital representation of the amplitude of the signal at each sampling point is transmitted along the telecommunications link.
  • This type of coding is pulse code modulation (PCM).
  • PCM pulse code modulation
  • the quality of sound reproduction using PCM depends on the sampling rate and also on the number of digits used to transmit each sample which determines the "quantization" or number of discrete amplitude levels that can be distinguished.
  • GSM European mobile radio standard
  • CELP code excited linear predictive coding
  • the technical problems addressed by the present invention are therefore to provide a method of speech coding and a speech coder which are capable of operation in real time without requiring excessive amounts of processing power, and to provide improved speech quality relative to that available from existing coders below 7Kb/s.
  • One of the objects of the present invention is to produce a speech coder in which an encoder and a decoder can be implemented on a single commercially available digital signal processing chip.
  • the encoders and decoders in accordance with the invention which will be described are each capable of implementation in real time using a DSP-32 floating point chip as manufactured by AT&T.
  • a second object of the present invention is to improve the digital speech quality significantly below 7Kb/s and to produce good quality at around 4Kb/s.
  • a speech coder for encoding an input speech signal for transmission over a digital channel of a telecommunications link, comprising means for sampling the input speech signal to produce output digital samples, means for dividing these digital samples into frames each consisting of a predetermined number of samples, a linear predictive filter for inverse filtering each frame and producing an output LPC residual signal for said frame comprising a said predetermined number of digital samples, and LPC parameters for said frame, and baseband extraction means, the speech coder being characterised by provision of down-sampling means for extracting from the output of said base-band extraction means d interleaved sequences, means for selecting one of said sequences which contains the maximum energy content and producing an output index representing the selected sequence, means for deriving pitch period and pitch gain indices from the selected sequence, means for removing long term correlation from the selected sequence to produce a remainder sequence, means for comparing the remainder sequence with an identifiable reference sequence, and for deriving a scale factor from the compared sequences, the scale
  • the baseband extraction means may comprise a weighting filter which amplifies a low frequency pitch component of the LPC residual signal and reduces the amplitude of higher frequency components of the LPC residual signal and said samples of the LPC residual signals are divided into blocks before being passed through the filter or alternatively the baseband extraction means may effect multipulse linear predictive analysis-by-synthesis whereby said baseband is obtained by minimizing an error between input speech signals and artificially reconstructed speech signals.
  • the means for comparing may include a vector quantizer for matching the remainder sequence with the most closely resembling one of a plurality of vectors stored in a codebook, each stored vector being a ramdom sequence having a Gaussian distribution and being identifiable by a unique index, and the output frame includes data representing the index of the selected vector.
  • the amount of data to be transmitted for each frame can be controlled in such a way as to select a bit rate in the range 2.4 Kb/s to 9.6 Kb/s which produces acceptable speech quality.
  • LPC filtering is already known and for a fuller description of the technique the reader is referred to: "Linear Prediction: A tutorial review” by J. Makhoul in Proc. IEEE, Vol-63, Pages 561-580, 1975.
  • LSP Line spectrum pair
  • Speech data compression by F.K. Soong and B.H. Juang in ICASSP-84 pages 1.10.1 to 1.10.4.
  • a Gaussian codebook vector quantization technique is employed which allows even lower bit rates to be achieved without reduction in speech quality. This type of vector quantization may be used after the LPC parameters have been transformed into line spectral pairs.
  • Vector quantization is also a standard technique and reference may be made for example to: "Vector quantization" by R.M. Gray in IEEE ASSP Magazine, Vol-1 pp4-29, 1984.
  • Vector quantization using a Gaussian codebook and a scale factor as described more fully in the accompanying specific description is believed to be novel.
  • the advantage of the proposed configuration is in the use of an analysis-by-synthesis procedure based around a pitch synthesis filter to select the optimum sequence from the Gaussian codebook and to compute its optimum scale factor.
  • the weighting filter is preferably a digital finite impulse response filter which has a gain-frequency characteristic which places emphasis on the large pulses in the signal represented by the input samples representing the LPC residual signal. These pulses occur periodically at a frequency corresponding to the underlying pitch of the voice signal. While the amplitude of these large pulses is relatively increased, the amplitude of the higher frequency components which contain proportionately less information is reduced. A typical filter characteristic is shown in Fig. 2.
  • the purpose of the weighting filter is to produce near-optimal excitation pulses as in the Multi-Pulse LPC proposed by P. Kroon et al in "Regular pulse excitation - A novel approach to effective and efficient multi-pulse coding of speech" IEEE Trans, ASSP-34 pp 1054-1063, 1986.
  • a pitch filter is used to remove from the LPC residual signal a signal representing the pitch pulses.
  • the parameters of such a pitch filter are preferably set by analysing the original LPC residual signal and a pitch filter memory.
  • the pitch filter is placed in a feedback loop in which data for transmission over the telecommunications link is fed to a decoder which carries out the inverse operations of the described coding vector quantizer and decimation to reproduce an LPC excitation signal, which is fed via the pitch filter and subtracted from the actual LPC residual signal so as to enhance the effect of the weighting filter and place more emphasis on the base band component of the speech signal.
  • the weighting filter, down-sampling and vector quantization steps effectively result in a minimisation of the difference between the output of the vector quantizer and the input to the weighting filter.
  • the data to be transmitted along the telecommunications link includes pitch data relating to the pitch amplitude and period of the feedback pitch filter. Extra bits are required for the transmission of this information and if it is necessary to keep the transmission rate constant, the bit rate occupied by the data output from the vector quantizer operating on the selected down-sampled sequence can be commensurately reduced without any reduction in the speech quality because there is now less information in the signal subject to vector quantization.
  • This pitch filter is able to operate in this manner because of the gain frequency characteristic of the weighting filter and the equivalent results would not be produced if a plain low-pass filter were used instead.
  • a speech coder for decoding speech encoded with the encoder in accordance with said first aspect, the coder comprising means for separating from a received frame the scale factor index, and the data representing the LPC parameters, means for outputting into a pitch synthesis filter a sequence corresponding to said identifiable sequence scaled by said scale factor, an interpolator for receiving said output sequence and said selected sequence index and interpolating zeros at appropriate positions in order to produce an LPC excitation signal, and an LPC synthesis filter for receiving said excitation signal and said data representing the LPC parameters and for restoring therefrom a sequence of digital samples representing the input speech signal.
  • the outputting means may comprise an inverse vector quantizer including a codebook corresponding to the codebook in the vector quantizer of the encoder, and the inverse vector quantizer receives said unique index and the scale factor index and outputs in response thereto, as the output sequence, a corresponding sequence scaled by said scale factor.
  • encoders and decoders described hereinafter are implemented as software instructions carried out in a digital signal processor such as the DSP-32 chip referred to previously.
  • the blocks shown in the drawings are intended merely to facilitate explanation of the functions of each of the processing steps carried out, rather than to indicate discrete components in the speech coder.
  • a speech channel of a telecommunications link using a speech coder requires an encoder at the voice signal input end and a decoder at the reception end. Therefore the speech coder associated with one end of the telecommunications link requires both an encoder and a decoder, which may be connected to separate channels in the case of a duplex link or the same channel in the case of a simplex link.
  • the encoder is diagrammatically illustrated in Figure 1 and the corresponding decoder is shown in Figure 3. Both the encoder and decoder may be implemented using the same digital signal processor.
  • the analogue speech signal input on line 2 of the encoder has a complex waveform (W) exhibiting, inter alia, relatively large amplitude pulses P, known as "pitch pulses", which are a characteristic of analogue speech signals.
  • the analogue speech signal is input on line 2 to a speech sampler 4 which samples the analogue speech signal and produces a series of digital samples.
  • a speech sampler 4 which samples the analogue speech signal and produces a series of digital samples.
  • the output samples are divided into frames of, for example, 200 8-bit samples each, and the encoder is effective to translate the samples in each frame into a number of quantization indices which represent the input waveform but consist of relatively few bits, thereby facilitating a low bit rate.
  • the frame size may be adjusted to suit the final bit rate required.
  • LPC linear predictive filter
  • the LPC parameters b j are computed in a processing circuit 12 for each input digital sample a i .
  • the LPC parameters b j derived for all the samples in the current frame are then fed to a parameter quantizer 16 which generates quantization indices therefrom, and these indices are routed on line 10 to a frame-forming circuit 44.
  • the quantization indices are also routed to a inverse quantizer 15 which re-generates the parameters b j , though the original and the re-generated parameters b j will not be identical due to the effect of processing the signals in the quantizer 16 and the inverse quantizer 15.
  • the re-generated LPC parameters b j are passed to an LPC inverse filter 14 which generates a further sample c i representing the difference between the corresponding input sample a i and a predicted value thereof, evaluated using the re-generated parameters b j .
  • the samples c i constitute an LPC residual signal, there being as many samples c i as there are input samples a i .
  • the LPC residual signal generated at the output of linear predictive filter 6 is then subjected to further quantization, as will now be described.
  • Each frame of samples of the LPC residual signal is divided into blocks.
  • the frame represented in Figure 8 has been divided into four blocks which, in this example, would each contain 50 samples.
  • each of these blocks is then fed separately into a weighting filter which is part of a processing circuit 18, shown in Figure 1.
  • the weighting filter is a finite impulse response digital filter with, for example, 11 taps.
  • the coefficients of the filter are such as to define a frequency-gain characteristic as shown in Figure 2, which is basically a low pass filter characteristic, but has important distinctions. As illustrated, low frequencies (below about 1 kHz) are subject to a positive gain which decays rapidly beyond 1 kHz. The purpose of this characteristic is to emphasise the relatively low frequency, periodic pulses of the voice signal which contain the most information and to diminish the significance of the higher frequency, intermediate parts of the signal much of which represents noise.
  • the blocks are each filtered separately. For an 11 tap filter to which the samples of successive blocks are fed continuously, the first five output samples and the last five output samples must be discarded. Therefore the number of output samples in the filtered block corresponds to the number of output samples in the input block.
  • the output samples from the filter for each block are then down-sampled by a decimation factor d in order to produce d decimated (interleaved) sequences.
  • Typical values of d are 3 or 4 though higher values may be used for lower bit rate channels.
  • the decimation factor is also partially determined by the block size, since each decimated sequence should be of equal length.
  • Processing block 18 is also effective to select one of the decimated sequences of each block by comparing the total energy contents of the sequences and selecting the sequence having the maximum energy. The energy of a sequence is determined by summing the squares of each of its constituent samples. An index s identifies the selected sequence and this index is also passed to the frame-forming circuit 44 on line 20.
  • the selected sequence is fed from processing block 18 to a vector quantizer 22 which is illustrated in more detail in Figure 4.
  • the concept of vector quantization is not novel per se but the particular characteristics of the vector quantizer which will now be described are considered to be unique in the present combination.
  • the input on line 24 to the vector quantizer is a signal consisting of a series of the selected decimated sequences derived from successive blocks.
  • the input signal for the current block i.e. the selected decimated sequence for that block
  • the existing contents of the memory 27 of a pitch synthesis filter 28 are fed to a control processor 26.
  • memory 27 contains a data sequence derived from the selected sequences of the immediately preceding frame
  • control processor 26 compares the sequences input thereto to obtain pitch indices p , h representing the pitch period and the pitch gain respectively of the decimated sequence in the current block.
  • the pitch period index p represents the number of shifts (relative to a datum position) that have to be performed in order to reach the position of maximum correlation of the current decimated sequence input on line 24 and the sequence stored in memory 27, and this shift usually represents the time interval between neighbouring pitch pulses P (shown in the waveform W).
  • the sequence selected from each remaining block in a frame could be correlated with the stored sequence at only eight (i.e. 23) different relative positions, distributed to either side of the pitch pulse already located by analysis of the sequence selected from the first block of the frame.
  • the pitch gain (represented by index h ) is calculated as the ratio of the cross-correlation of the selected input sequence and the pitch filter memory (at the position of maximum correlation) normalised with respect to the block energy of the contents of the pitch filter memory.
  • the pitch indices p , h generated in this manner are output from the control processor 26 to the pitch synthesis filter 28 and also to the frame-forming circuit 44.
  • this sequence is then subjected to vector quantization. This involves comparing the pattern of the current, selected sequence with the pattern of each of a number of reference sequences or vectors stored in a Gaussian codebook 38 in order to determine which of these reference sequences it most closely resembles, the selected reference sequence being represented by a unique index f .
  • pitch synthesis filter 28 (which would detract from the effectiveness of the matching procedure) is subtracted from the current sequence input on line 24.
  • the contents of the memory 27 of the pitch synthesis filter 28 are transferred on line 25 to the memory 31 of an otherwise identical auxiliary pitch synthesis filter 29 which is used to compute the pitch synthesis filter memory response.
  • the pitch synthesis filter 29, which is an infinite impulse response filter, is clocked with a zero input to find its memory response which is then output on line 33.
  • This memory response is fed into a subtractor 35 together with the current input sequence on line 24, thereby to produce a difference or reference signal on line 37 at the output of the subtractor.
  • This setting of the pitch synthesis filter 28 is carried out initially for each block to be processed.
  • the zero input pitch filter response is subtracted from the input signal in order to reduce the mean-squared error during the subsequent matching operation which is designed to identify which one of a plurality of vectors stored in Gaussian codebook 38 most closely matches the input signal on line 37.
  • the pitch synthesis filter 28 is fed with different sequences from the Gaussian codebook 38. These sequences, together with the pitch data, are used to generate output signals which are routed to a further subtractor 32 on line 30, the other input to subtractor 32 being the difference, or reference signal on line 37.
  • the output from subtractor 32 is therefore a difference signal representing the mismatch between the two inputs to the subtractor.
  • This mismatch or "error" for each successive vector input to the pitch synthesis filter 28 is computed by summing the squares of the sample values in an error computing processor 34. The error for each successive signal is fed back to the control processor 26 and the error processor also produces an output signal on line 36 to indicate that the error computation is complete for that input signal.
  • the number of different pattern sequences or vectors stored in Gaussian codebook 38 determines the accuracy of quantization.
  • the purpose of the vector quantizer is to determine which of the vectors stored in the codebook most closely resembles the pattern (but not necessarily the magnitude) of the selected decimated sequence which is input to the vector quantizer. Once the closest vector from the Gaussian codebook 38 has been identified by the vector quantizer, the entire decimated sequence can be represented by the index f of this vector, the analysed pitch characteristics h , p and a scale factor g , the derivation of which is now described.
  • Each of the vectors in Gaussian codebook 38 is a random sequence which has a zero mean Gaussian energy distribution and a normalized energy content. Because of this, each signal output from the Gaussian codebook is multiplied, in a multiplication circuit 40 by the optimal scale factor g which is computed by control processor 26 from the energy contents of the signals on lines 37 and 30.
  • the optimal scale factor g is given by the cross-correlation of the signals on lines 37 and 30 divided by the energy of the signal on line 30.
  • the signal on line 30 is first computed with g set at 1. An aim of the scale factor calculation is to reduce the scale factor towards zero if the input sequence mainly contains noise and there is no significant correlation between the signal on line 37 and the signal on line 30.
  • the energy in the input sequence on line 37 fluctuates over a relatively wide range, and so as many as 5 or 6 bits may be needed to define scale factor g .
  • the number of bits can be reduced significantly by further normalising the scale factor (before coding) with respect to the energy in pitch filter memory 28, and this leads to a further reduction in the bit rate of the speech coder.
  • the memory 27 of the pitch synthesis filter 28 is updated as follows in readiness for processing the next block.
  • the original contents of memory 27, which were transferred to memory 31, are returned to memory 27 and the selected vector from the Gaussian codebook 38 is scaled by the optimum scale factor g and input to the pitch synthesis filter 28.
  • the process of clocking the filter finishes (after clocking the required number of times depending on the number of elements in each vector in the Gaussian codebook) the resultant contents of the memory 27 are retained for processing the next block.
  • the output from the vector quantizer 22 is fed on line 42 to the frame forming circuit 44 which also receives inputs from the quantizer 16 of the linear predictive filter 6 and from the decimation processor 18.
  • the frame forming circuit assembles this input data into a predetermined standard format for transmission over the channel 46.
  • the index f which represents the selected vector from the Gaussian codebook 38, may consist of as many as 8 or 9 bits, depending on the accuracy of the quantization.
  • the reference pattern is derived from the memory response of the auxiliary pitch synthesis filter 29, and the corresponding index may then be defined using fewer bits, leading to a reduced bit rate.
  • the reference pattern is derived from the memory response of pitch filter 29 by means of an alternative circuit, shown generally at 47, which is used in place of the Gaussian codebook 38.
  • the reference pattern is generated by suitably clipping the memory response using a clipping circuit 48.
  • the decoder illustrated in Figure 3 essentially carries out the inverse of each of the operations carried out in the described encoder working on the basis of the data transmitted over the channel 46.
  • This data is first fed to a frame decoder 48 which extracts the various items of data transmitted.
  • the data generated by the vector quantizer 22 in the encoder are fed to an inverse vector quantizer 50 which contains a memory storing identical vectors to those stored in the Gaussian codebook 38.
  • the index f generated by the encoder determines which of these stored sequences is read out and multiplied by the scale factor g . If circuit 47 is used in the encoder instead of the Gaussian codebook 38, then the memory of the inverse quantizer 50 would contain a corresponding vector derived from the memory response of the pitch filter.
  • the pitch parameters h and p are used to control a pitch synthesis filter corresponding to filter 28 of the vector quantizer in the encoder, and this adds in the pitch pulse components. Therefore the output of the inverse vector quantizer 50 is a representation of the decimated sequence which was fed to the vector quantizer in the encoder. Zeros must be interpolated into this decimated sequence in order to produce an LPC excitation signal which corresponds to a representation of the LPC residual signal.
  • the frame decoder supplies to an interpolation processor 52 the sequence index s so that d - 1 zeros may be interpolated between successive samples of the sequence and an appropriate number of zeros interpolated at the beginning and end in order to place the samples of the decimated sequence in their correct positions in the excitation signal.
  • the output of the interpolation processor 52 is then fed to an LPC synthesis filter 54.
  • This filter receives control inputs from the frame decoder representing the quantized parameters and, in a known manner, uses the excitation signal to produce a representation of the original digital samples. Since the LPC parameters have been quantized, an inverse quantization must first be carried out (as occurred in circuit 15 of the encoder in Figure 1), before the LPC synthesis filter can operate to restore the original speech samples. These samples are then fed out via a digital-to-analogue converter 56 to reproduce an analogue signal corresponding to the voice signal originally input on line 2.
  • the LPC synthesis filter 54 in the decoder does not have an "ideal" memory response, and this tends to detract from the quality of the processed signals.
  • the memory response of filter 54 may be subtracted from the digital samples at the output of sampler 4 of the encoder, before these samples are fed to the linear predictive inverse filter 6.
  • the LPC synthesis filter 54 is clocked from time-to-time (once per frame, say) with a zero input, and the zero input memory response of the filter is passed to the memory 80 of an identical LPC synthesis filter 81, the output of which is connected to one input of a subtraction circuit 82 which interconnects the speech sampler 4 and the coder 6 of the encoder ( Figure 1).
  • Filter 81 is also clocked with a zero input and by this means the zero input memory response of filter 81, which is the same as that of filter 54 in the decoder, is subtracted by subtraction circuit 82 from the input digital samples produced by the speech sampler 4.
  • the encoder shown in Figure 5 is similar to that illustrated in Figure 1. However, in this embodiment an additional feedback loop comprising a decoder and interpolation processor 60 and a pitch filter 62 is included.
  • the output of the pitch filter 62 is fed to a subtractor 64 connected to receive the LPC residual signal from the inverse 14 filter of the linear predictor 6.
  • the decoder and interpolation processor generates from the output frame for transmission a representation of the excitation signal which in the decoder proper would be fed to the LPC synthesis filter.
  • this feedback loop it is fed to the pitch filter 62 which removes from it the pitch pulses so that the output of subtractor 64 which is fed to the weighting filter and decimation processor 18 contains a less significant contribution from the pitch pulses.
  • the parameters of the pitch filter 62 that is the pitch gain q and pitch period r are determined by analysing the LPC residual signal output from filter 14 in a processor not shown.
  • the contents of a memory of the pitch filter may also be used in the analysis.
  • This pitch data must also be transmitted over the channel 46 and is therefore also supplied to the frame former 44.
  • the pitch data q , r may be updated for each block, each frame or even less frequently depending on the bit rate constraints. However, because there is less information now present in the signal which is fed to the vector quantizer, the number of bits required for transmitting its output data may be reduced. In particular, the size of the Gaussian codebook may be restricted.
  • the decoder for use in conjunction with the encoder of Figure 5 is illustrated diagrammatically in Figure 6. It is essentially identical to the decoder described with reference to Figure 3 except that the output from the interpolator 52 is not fed directly to the LPC synthesis filter but is first fed through a pitch synthesis filter 68 which is controlled by the pitch gain q and the pitch period r transmitted over the channel 46.
  • This pitch synthesis filter restores to the excitation signal a series of pitch pulses corresponding to those identified in the originally encoded LPC residual signal. In this way the interpolated zeros are to some extent overwritten by contributions from the pitch synthesis filter. This results in much smoother quality, because the high frequency distortion created by the spectral folding of the interpolation is largely eliminated.
  • the input on line 68 to this quantizer is the parameters after having been transformed in a processor (not shown) into the line spectral pairs domain using transforms as described in the literature. A transformation into some other domain which will enable correlation or matching of similarities between the transformed LPC parameter vectors and vectors stored in a codebook may also be used.
  • a new input sequence is generated for each frame of input digital samples.
  • Each input sequence is fed to a control processor 70 and to a first codebook 72.
  • the control processor 70 analyses an error signal on line 74 produced after matching in the first codebook and with every sequence or vector of a second codebook 76 to compute a scale factor w .
  • the codebook 72 is a first in first out (FIFO) memory store.
  • the vectors generated for storage in this codebook are derived from previously received sequences as will be described in more detail later.
  • the input sequence is matched with each of the stored vectors in turn and, by a process of least squares minimisation conducted by the control processor, the index t of the vector which most closely matches the input sequence is generated.
  • the vector V t so identified is output from the codebook 72 to a subtractor 74 which also receives the actual input sequence.
  • the output from the subtractor 74 is therefore an error signal representing the difference between the input sequence and the most closely matched vector of the codebook 72.
  • This error signal is fed to the control processor and to a second Gaussian codebook 76.
  • the vectors stored in the codebook 76 are each normalised random sequences with a Gaussian distribution and zero mean. This codebook 76 is therefore of the same type as the codebook 38 used in the previously described vector quantizer of Figure 4.
  • the index u of the vector in the codebook 76 which most closely matches the input error signal is determined by a least squares minimisation technique.
  • the selected vector V u is then output to a multiplier 78 where it is multiplied with the optimal scale factor w producing a sequence w V u which is added in an adder 79 to the vector V t selected from the first codebook 72.
  • the output of the adder 79 is then placed into the codebook 72 displacing the oldest previously stored vector.
  • the contents of the codebook 72 are therefore being continuously updated so that the effectiveness of this codebook continuously increases while the input voice signal has characteristics of the same speaker.
  • the outputs for transmission over the channel 46 link are the two indices t , u and the optimal scale factor w .

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Signal Processing (AREA)
  • Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Human Computer Interaction (AREA)
  • Acoustics & Sound (AREA)
  • Multimedia (AREA)
  • Compression, Expansion, Code Conversion, And Decoders (AREA)
  • Spectroscopy & Molecular Physics (AREA)
  • Transmission Systems Not Characterized By The Medium Used For Transmission (AREA)
EP89302481A 1988-03-16 1989-03-14 Sprachcodierung Ceased EP0333425A3 (de)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
GB8806185 1988-03-16
GB8806185A GB8806185D0 (en) 1988-03-16 1988-03-16 Speech coding

Publications (2)

Publication Number Publication Date
EP0333425A2 true EP0333425A2 (de) 1989-09-20
EP0333425A3 EP0333425A3 (de) 1990-02-07

Family

ID=10633500

Family Applications (1)

Application Number Title Priority Date Filing Date
EP89302481A Ceased EP0333425A3 (de) 1988-03-16 1989-03-14 Sprachcodierung

Country Status (2)

Country Link
EP (1) EP0333425A3 (de)
GB (1) GB8806185D0 (de)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP0501421A2 (de) * 1991-02-26 1992-09-02 Nec Corporation Sprachkodiersystem
ES2042410A2 (es) * 1992-04-15 1993-12-01 Control Sys S A Metodo de codificacion y codificador de voz para equipos y sistemas de comunicacion.
US5946651A (en) * 1995-06-16 1999-08-31 Nokia Mobile Phones Speech synthesizer employing post-processing for enhancing the quality of the synthesized speech

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP0125423A1 (de) * 1983-04-13 1984-11-21 Texas Instruments Incorporated Vocoder unter Anwendung von Grundfrequenzermittlung auf gefiltertem LPC-Differenzsignal beruhend

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP0125423A1 (de) * 1983-04-13 1984-11-21 Texas Instruments Incorporated Vocoder unter Anwendung von Grundfrequenzermittlung auf gefiltertem LPC-Differenzsignal beruhend

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
ELECTRONICS LETTERS, vol. 23, no. 24, 19th November 1987, pages 1286-1288, Hitchin, GB; A. KONDOZ et al.: "Vector-quantised transform coder for speech coding at 9.6kbit/s and below" *
PROCEEDINGS ICASSP 87, INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, Dallas, Texas, 6th-9th April 1987, vol. 3, pages 1637-1640, IEEE, New York, US; R.C. ROSE et al.: "Quality comparison of low complexity 4800 bps self excited and code excited vocoders" *

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP0501421A2 (de) * 1991-02-26 1992-09-02 Nec Corporation Sprachkodiersystem
EP0501421A3 (en) * 1991-02-26 1993-03-31 Nec Corporation Speech coding system
ES2042410A2 (es) * 1992-04-15 1993-12-01 Control Sys S A Metodo de codificacion y codificador de voz para equipos y sistemas de comunicacion.
US5946651A (en) * 1995-06-16 1999-08-31 Nokia Mobile Phones Speech synthesizer employing post-processing for enhancing the quality of the synthesized speech
US6029128A (en) * 1995-06-16 2000-02-22 Nokia Mobile Phones Ltd. Speech synthesizer

Also Published As

Publication number Publication date
GB8806185D0 (en) 1988-04-13
EP0333425A3 (de) 1990-02-07

Similar Documents

Publication Publication Date Title
EP0331857B1 (de) Verfahren und Einrichtung zur Sprachkodierung mit niedriger Datenrate
US5371853A (en) Method and system for CELP speech coding and codebook for use therewith
US5265190A (en) CELP vocoder with efficient adaptive codebook search
EP0573216B1 (de) CELP-Vocoder
US5067158A (en) Linear predictive residual representation via non-iterative spectral reconstruction
US6401062B1 (en) Apparatus for encoding and apparatus for decoding speech and musical signals
EP0331858B1 (de) Verfahren und Einrichtung zur Sprachkodierung mit mehreren Datenraten
US5359696A (en) Digital speech coder having improved sub-sample resolution long-term predictor
US6078880A (en) Speech coding system and method including voicing cut off frequency analyzer
US4704730A (en) Multi-state speech encoder and decoder
EP0751494B1 (de) System zur sprachkodierung
US6119082A (en) Speech coding system and method including harmonic generator having an adaptive phase off-setter
US6081776A (en) Speech coding system and method including adaptive finite impulse response filter
US6023672A (en) Speech coder
US6138092A (en) CELP speech synthesizer with epoch-adaptive harmonic generator for pitch harmonics below voicing cutoff frequency
JPH0395600A (ja) 音声コーディング装置及び音声エンコーディング方法
EP0450064B1 (de) Numerischer sprachkodierer mit verbesserter langzeitvorhersage durch subabtastauflösung
US4945565A (en) Low bit-rate pattern encoding and decoding with a reduced number of excitation pulses
CA1144650A (en) Predictive signal coding with partitioned quantization
US5873060A (en) Signal coder for wide-band signals
US5649051A (en) Constant data rate speech encoder for limited bandwidth path
JPS5887936A (ja) デジタル情報伝送方式
Chen et al. Vector adaptive predictive coding of speech at 9.6 kb/s
Gersho et al. Fully vector-quantized subband coding with adaptive codebook allocation
EP0333425A2 (de) Sprachcodierung

Legal Events

Date Code Title Description
PUAI Public reference made under article 153(3) epc to a published international application that has entered the european phase

Free format text: ORIGINAL CODE: 0009012

AK Designated contracting states

Kind code of ref document: A2

Designated state(s): AT BE CH DE ES FR GB GR IT LI LU NL SE

PUAL Search report despatched

Free format text: ORIGINAL CODE: 0009013

AK Designated contracting states

Kind code of ref document: A3

Designated state(s): AT BE CH DE ES FR GB GR IT LI LU NL SE

17P Request for examination filed

Effective date: 19900427

17Q First examination report despatched

Effective date: 19911028

STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: THE APPLICATION HAS BEEN REFUSED

18R Application refused

Effective date: 19930923