EP0075311B1 - Arrangement for speech transmission based on the channel vocoder principle - Google Patents

Abstract

A synthesizer for a channel vocoder for the transmission of speech with considerable frequency band reduction in digital technology provides a reduction of circuit expense. This is accomplished, given non-recursive filters having finite impulse response, not with a weighting of the pulse-shaped excitation variable with the transmitted envelope values of the spectral channels, but with a time variance of the filters by weighting their filter coefficients with the transmitted envelope values. In addition, with proper dimensioning of the filter coefficients, an optimum speech quality of synthesized speech is obtained at the output of the synthesizer even given elimination of the transmission of a voiced/voiceless signal.

Description

The invention relates to an arrangement for the transmission of speech according to the channel vocoder principle, in which, in a vocoder analyzer on the transmission side, envelope values of the speech signal to be transmitted for a number of M spectral channels predetermined by a filter bank, as well as additional speech-specific parameters, such as basic speech frequency and width, differ by frequency position and width Vocal characteristics derived and combined in a suitable manner to form a digital sum signal in the rhythm of successive frames each comprising an analysis interval are transmitted to the reception side, in which furthermore the digital sum signal on the reception side is again divided into the individual M spectral channels and, if appropriate, the channels assigned to the additional language-specific parameters and these channel signals are then fed to the receiver-side vocoder synthesizer, which in turn corresponds to the transmitter-side vocoder analyzer has a computing filter bank and a pulse generator and outputs the generated synthetic speech signal on the output side.

Channel vocoder transmission methods are known, for example, from the literature IEEE Transactions on Audio and Electroacoustics, Vol. AU-15, No. 4, Dec. 1967, pages 148 to 161. As a rule, the same filter bank is used for such channel vocoders in the analysis section and in the synthesis section. In the case of voice transmission in half duplex, this results in an inexpensive outlay because the filter bank can be switched over and thus only needs to be present once per device.

The transition from analog circuits to digital circuits, which is taking place more and more in the course of advancing integrated technology, also leads to channel vocoders being implemented exclusively in digital technology. As the literature reference IEEE Transactions on Acoustics, Speech, and Signal Processing, Vol. ASSP-29, No. 1 Feb. 1981, pages 13 to 23, in particular page 16, right column, first paragraph shows, are used for the implementation of the filter bank mentioned in digital version IIR filter and FIR filter. However, the effort for such filters is considerable. This applies in particular to the non-recursive filter with a finite impulse response (FIR fliter), because this must be of a high atomic number. The digital recursive filter with an infinite impulse response (IIR filter) can be implemented with a significantly lower atomic number for the purpose mentioned, but does not have the favorable properties of the FIR filter.

So-called “switched capacitor filters” can also be used as digital filters, which are available in an integrated embodiment and manage with a small installation volume. However, like all sampling filters, such filters have the disadvantage that their transfer function is repeated periodically in the frequency domain. In other words, the input signal for such a filter must not have any spectral components above half the sampling frequency. As practice shows, the upper band limit of the input signal must be further away from half the sampling frequency if interference is to be avoided. With regard to the analysis part, there are no problems with this, since this distance can be reliably maintained with simple means. In the synthesis part, on the other hand, this requirement cannot be guaranteed without considerable additional expenditure due to the pulse-shaped excitation function. In other words, considerable analog screening means have to be used here in order to sufficiently suppress disruptive effects in the form of a chirping background noise.

The invention has for its object to provide a solution for a digital filter bank for the synthesis part of a channel vocoder implemented in digital technology, which can ensure optimal speech quality of the generated synthetic speech with relatively little technical effort.

Starting from an arrangement for the transmission of speech according to the channel vocoder principle of the type described in the introduction, this object is achieved according to the invention in that the filter bank designed in digital technology consists of non-recursive time-variant filters with a finite impulse response (FIR filter) and that the filter bank the excitation variable of the pulse generator is supplied on the input side with a constant pulse amplitude and that the time variance of the gain factor of the filter bank is brought about in multipliers by multiplying the filter bank coefficients by the envelope values transmitted in rhythm with successive frames.

The invention is based on the essential finding that the impulse response of a FIR filter amounts to the same whether the pulse train supplied to its input is weighted with the transmitted envelope values or the weighting is carried out by multiplying the filter coefficients by the transmitted envelope values . The inventive design of the filter bank as a time-variant filter bank, to the input side of which the excitation variable of the pulse generator is supplied with a constant amplitude, reduces in an extremely advantageous manner the multiplications to be carried out in the synthesis part and thus the filter effort.

In a first preferred embodiment, the FIR filters, to which the excitation variable is fed on the input side, each have a chain connection of N-1 identical delay stages with a maximum of N taps. Each FIR filter has a number of switches corresponding to the taps, the control inputs of which are connected to the taps. Each switch connects the output of a multiplier to the filter output. The inputs of the multipliers together form the Input for the assigned transmitted envelope value and at each of the other inputs of the multipliers there is a filter coefficient, which in their entirety determine the filter response.

In a second preferred arrangement, the filter bank is an FIR sum filter, which has a chain connection of N-1 equal delay stages with a maximum of N taps and has a number of switches corresponding to the taps, the control inputs of which are connected to the taps. Each switch connects the output of a multiplier summing circuit to the filter output, which carries out the product formation of the individual filter coefficients with the associated transmitted envelope values and forms the product sum.

The usually for reasons of redundancy reduction exclusively in powers of two transmitted envelope values give the possibility of multiplying them with the filter coefficients in a sliding mechanism, which leads to a further considerable reduction in expenditure.

In a further development of the invention, the excitation quantity supplied to the FIR filter or the FIR sum filter can be a pulse sequence that is generated periodically by a pulse generator and can be controlled in its repetition rate by the transmitted basic speech frequency value. In this context, the filter coefficients are dimensioned such that an optimal speech quality of the synthesized speech is guaranteed.

A first such preferred dimensioning can consist in that the filter coefficients of the FIR filter or the filter areas of the FIR sum filter with center frequencies <2 kHz for impulse responses and those of the FIR filters or the filter areas of the sum filter with center frequencies> 2 kHz for noise responses are measured.

A further preferred dimensioning can consist in dimensioning all filter coefficients of the FIR filter or the FIR sum filter for noise responses.

• Figur 1 das Blockschaltbild des Syntheseteils eines üblichen Kanalvocoders,
• Figur 2 das Fig. 1 entsprechende Blockschaltbild eines erfindungsgemäßen Syntheseteils,
• Figur 3 das Blockschaltbild eines digitalen Bandfilters des Syntheseteils nach Fig. 2,
• Figur 4 ein weiteres Syntheseteil mit einem digitalen Summenfilter nach der Erfindung,
• Figur 5 eine weitere Ausführungsform eines Syntheseteils nach der Erfindung in schematischer Darstellung.

The invention will be explained in more detail below with reference to exemplary embodiments shown in the drawing. Mean in the drawing

• FIG. 1 shows the block diagram of the synthesis part of a conventional channel vocoder,
• 2 shows the block diagram corresponding to FIG. 1 of a synthetic part according to the invention,
• FIG. 3 shows the block diagram of a digital band filter of the synthesis part according to FIG. 2,
• FIG. 4 shows a further synthesis part with a digital sum filter according to the invention,
• Figure 5 shows a further embodiment of a synthetic part according to the invention in a schematic representation.

The envelope values Ai combined within an analysis interval on the transmission side to form a time-division multiplex frame, as well as the fundamental frequency information No and the voiced-unvoiced criterion Sc are first of all returned to the individual channels, namely the channels, on the receiving side in the demultiplexer DEMUX, to which the sum signal ss is fed on the input side M spectral channels for the M envelope values A1, A2 ... AM as well as the channel for the voiced-unvoiced criterion Sc and the channel for the fundamental frequency information No are divided.

The actual synthesis part ST has the filter bank with the bandpasses B1, B2 ... BM, to which the envelope values are fed via a modulator M1, M2 ... MM. Depending on the voiced-unvoiced criterion S, either the excitation variable x (n) in the form of the periodic pulse sequence generated by the pulse generator PB or in the form of the pseudo-random pulse sequence generated by a random pulse generator PNG is fed to the second input of the modulators. The pulses of the pulse generator PG or of the random pulse generator PNG which are output at the modulator outputs and weighted by the envelope values A1, A2 ... AM are passed over the bandpass filters B1, B2 ... BM. The filter responses Y1 (n), Y2 (n) ... YM (n) are then combined via the summer SU to the synthesized speech signal y (n) at the output of the synthesis part ST.

For the execution of the band filter B1, B2 ... BM as FIR filter, the output signal y (n) is obtained taking into account the filter coefficients h1 (k), h2 (k) ... hM (k ) the relationship for these bandpass filters

It has been assumed that each of the band filters B1, B2 ... BM N representing a FIR filter has taps and that the same time interval of two successive pulses is chosen from N-1 identical delay elements with the delay time. The expression in equation (1) after the first sum symbol represents the respective filter response yi (n), so that equation (1) can also be written

As can be seen from equation (1), the same result is obtained for the output signal y (n) if, instead of a weighting, that of the pulse generator DG or of the random pulse generator PNG supplied excitation variable x (n) with the envelope values Ai, the filter coefficients hi (k) are multiplied with the envelope values Ai. The equation (1) for the output signal of the synthesis part y (n) can therefore be rewritten as follows.

If one sets for the product Ai - 'hi (k) = gi (k), the result is -

The expression after the first sum symbol in turn represents the output function yi (n) of a single filter, so that it also applies here

The synthesis part ST which realizes equations (3) to (5) and is constructed with FIR filters is shown in FIG. 2. Instead of the time-inverted bandpasses B1, B2 ... BM according to FIG. 1, the bandpasses B1 ', B2', ... BM 'representing time-variant FIR filters are now used. The excitation variable x (n) is fed directly to the inputs of these bandpass filters in the form of a pulse train with a constant pulse amplitude. The time variance of the filters is controlled via the envelope values A1, A2 ... AM by weighting the filter coefficients hi (k) by forming the products gi (k). This weighting must be done once for each incoming frame of the sum signal ss for each filter.

FIG. 3 shows a bandpass filter Bi 'in the form of such a time-variant FIR filter. It has N-1 identical delay stages Z- ¹ and N taps connected in chain. The N taps are fed to control inputs of N switches S, which each connect the output of a multiplier MU to the sum line 1 supplying the filter response yi (n). Each multiplier MU assigned to a switch S is supplied with the associated envelope value Ai at its one input. The filter coefficients hi (0), hi (1), ... hi (N-1) are present at the N second inputs of the N multipliers MU. The excitation variable x (n) supplied to the input of the chain of delay elements has only the function of controlling the switches S "on or" within the filter, depending on whether no pulse or a pulse occurs. This has become possible because the excitation variable itself is no longer weighted by the envelope values. The transition from time-invariant to time-variant FIR filters thus saves M multipliers. As has already been mentioned, these multipliers can be designed in an extremely simple manner as sliding mechanisms, since the transmitted envelope values can only have values which are either zero or potencies of two for reasons of redundancy reduction.

an die Summenleitung 1 angeschaltet und auf diese Weise das Ausgangssignal y(n) am Ausgang des Syntheseteils ST' gewonnen.FIG. 4 shows an embodiment of the synthesis part ST according to FIG. 2, in which the bandpasses B1 ', B2' ... BM 'are combined to form a sum filter SB. The sum filter SB of the synthesis part designated ST in FIG. 4, in accordance with FIG. 3, in turn has the chain connection of N same delay stages Z- ¹ with N taps, each of which is supplied to the control input of one of N switches. In contrast to the individual filter according to FIG. 3, instead of a multiplier, a multiplier-summing arrangement MS is provided here, each of which has M multipliers. The outputs of the M multipliers are connected to the associated switch S via a summer SU. The filter coefficients hi (k) of all equivalent taps of the individual filters are supplied to one input of the M multipliers MU of a multiplier-summing arrangement MS, while the envelope values A1, A2 ... AM are present at their second inputs. Depending on the fact that a pulse of the excitation variable x (n) is present at the control input of a switch or not, the respective sum function

connected to the sum line 1 and in this way the output signal y (n) obtained at the output of the synthesis part ST '.

As further studies on which the invention is based have shown, the transmission of the voiced-unvoiced criterion Sc and the random generator PNG can be dispensed with in an extremely advantageous manner in such a synthesis part with time-variant FIR filters with a suitable measurement of the filter coefficients defining the impulse response or the noise response will. This is shown schematically in Fig. 5. The filter coefficients hi '(k) are expediently determined here experimentally so that the synthesized speech signal has an optimal speech quality at the output. As has already been explained, this can be done in such a way that the filter coefficients hi '(n) of the band-pass filters B1', B2 '... BM' or the filter areas of the sum filter SB with center frequencies <2 kHz for impulse responses and those the bandpasses or the filter areas of the sum filter are dimensioned with center frequencies> 2 kHz for noise responses. However, it is also possible in an extremely advantageous manner to dimension all filter coefficients hi '(n) of the bandpass filter or the sum filter in a suitable manner for noise responses.

Claims (7)

1. An arrangement for transmitting speech in accordance with the channel vocoder principle, wherein, at the transmitting end, from the speech signal which is to be transmitted there are obtained envelope curve values for a number - predetermined by a filter bank - of M spectral channels which differ in respect of frequency position and frequency with and possibly additional speech-specific parameters, such as fundamental speech frequency and tuning characteristic, and these are transmitted in a suitable manner to the receiving end, combined to form a digital sum signal, in the timing of consecutive frames which each comprise one analysis interval, wherein moreover at the receiving end the digital-sum signal is re-divided between the individual M spectral channels and where appropriate between the channels assigned to the additional speech-specific parameters and these channel signals are then fed to the receiving end vocoder synthesiser which itself comprises a filter bank, corresponding to the transmitting-end vocoder analyser, and a pulse generator and at the output end emits the generated, synthetic speech signal, characterised in that the filter bank, which is designed in digital technology, comprises non-recursive, time-variant filters (B1' ... BM') having a finite pulse response (FIR-filters), that moreover the filter bank is supplied at its input end with the excitation value (x(n)) of the pulse generator (PG, PNG) with a constant pulse amplitude and that the time variants of the amplification factor of the filter bank is obtained by multiplying the filter bank coefficients (hi(n), hi'(n)) in multipliers (MU) by the envelope curve values (A1 ... AM) which are transmitted in the timing of consecutive frames.

2. An arrangement as claimed in claim 1, characterised in that the FIR-filters, which are supplied at their input end with the excitation value (x(n)), each comprise a chain circuit of N-1 identical delay stages (Z-¹) having a maximum of N tappings, that each FIR-filter also has a number - corresponding to the tappings - of switches (S) whose control inputs are connected to the tappings, that moreover each switch connects the output of a multiplier (MU) to the filter output and that the first inputs of the multipliers commonly form the input for the assigned, transmitted envelope curve value and a filter coefficient (hi(n)) occurs at each of the other inputs of the multipliers, which filter coefficients together determine the filter response.

3. An arrangement as claimed in claim 1, characterised in that the filter bank is a FIR-sum filter which includes a chain circuit of N-1 identical delay stages (Z-¹) having a maximum of N tappings, that moreover a number - corresponding to the tappings - of switches (S) is provided whose control inputs are connected to the tappings and that each switch connects of the filter output the output of a multiplier- adder circuit (MS) which forms the product of the individual filter coefficients (hi(n)) with the associated, transmitted envelope curve values (Ai) and forms the product sum.

4. An arrangement as claimed in one of the preceding claims, characterised in that the multipliers comprise shift devices.

5. An arrangement as claimed in one of the preceding claims, characterised in that, dispensing with unvoiced-voiced switch-over facilities, the excitation value (x(n)) which is fed to the FIR-filters (B1' ... BM', Bi') or the FIR-sum filter (SB) is a pulse sequence which is periodically produced by a pulse generator (PG) and whose repetition frequency can be controlled by the transmitted fundamental speech frequency value (No) and that the filter coefficients (hi(n)) are contrived to be such that an optimum speech quality of the synthesised speech is thereby ensured.

6. An arrangement as claimed in claim 5, characterised in that the filter coefficients (hi'(n)) of the FIR-filters (B1' ... BM', Bi) and of the filter ranges of the FIR-sum filter (SB) with middle frequencies of < 2 kHz are specified for pulse responses and the filter coefficients of the FIR-filters and the filter ranges of the FIR-sum filter with middle frequencies of > 2 kHz are specified for noise responses.

7. An arrangement as claimed in claim 5, characterised in that all the filter coefficients (hi'(n)) of the FIR-filters (B1' ... BM', Bi') and of the FIR-sum filter (SB) are specified for noise responses.

Images