Spectral magnitude modeling and quantization in a frequency domain interpolative speech codec system

• US 6,493,664 B1
• Filed: 04/04/2000
• Issued: 12/10/2002
• Est. Priority Date: 04/05/1999
• Status: Expired due to Term

• Current Assignee: Hughes Network Systems LLC (Echostar Corporation)
• Original Assignee: Hughes Electronics Corporation (Rtx Corporation)
• Inventors: Swaminathan, Kumar, Zakaria, Gaguk, Udaya Bhaskar, Bangalore R., Nandkumar, Srinivas
• Primary Examiner(s): Dorvil, Richemond
• Assistant Examiner(s): Nolan, Daniel A
• Application Number:
  US09/542,793
• Time in Patent Office: 980 Days
• Field of Search: 704/203-207, 704/218, 704/222, 704/265, 704/211, 704/221
• US Class Current: 704/222
• CPC Class Codes: G10L 19/005 Correction of errors induced by the transmission channel, if related to the coding algorithm
• CPC Class Codes: G10L 19/02 using spectral analysis, e.g. transform vocoders or subband vocoders
• CPC Class Codes: G10L 19/0204 using subband decomposition
• CPC Class Codes: G10L 19/04 using predictive techniques
• CPC Class Codes: G10L 19/083 the excitation function being an excitation gain G10L25/90 takes precedence
• CPC Class Codes: G10L 19/09 Long term prediction, i.e. removing periodical redundancies, e.g. by using adaptive codebook or pitch predictor
• CPC Class Codes: G10L 19/18 Vocoders using multiple modes
• CPC Class Codes: G10L 2019/0012 Smoothing of parameters of the decoder interpolation
• CPC Class Codes: G10L 2025/783 based on threshold decision
• CPC Class Codes: G10L 25/27 characterised by the analysis technique
• CPC Class Codes: G10L 25/30 using neural networks
• CPC Class Codes: G10L 25/78 Detection of presence or absence of voice signals switching of direction of transmission by voice frequency in two-way loud-speaking telephone systems H04M9/10
• CPC Class Codes: G10L 25/90 Pitch determination of speech signals

Abstract

Encoding of prototype waveform components applicable to telecommunication systems provides improved voice quality enabling a dual-channel mode of operation which permits more users to communicate over the same physical channel. A prototype word (PW) gain is vector quantized using a vector quantizer (VQ) that explicitly populates a codebook by representative steady state and transient vectors of PW gain for tracking the abrupt variations in speech levels during onsets and other non-stationary events, while maintaining the accuracy of the speech level during stationary conditions.

Claims

1. A frequency domain interpolative coding system for magnitude modeling and quantization of prototype waveforms, comprising:

a linear prediction (LP) front end responsive to an input signal providing LP parameters which are quantized and encoded over predetermined intervals and used to compute a LP residual signal;

an open loop pitch estimator responsive to said LP residual signal;

a pitch quantizer;

a pitch interpolator;

said open loop pitch estimator, said pitch quantizer, and said pitch interpolator yielding a pitch contour within the predetermined interval;

a signal processor responsive to said LP residual signal and the pitch contour for extracting a prototype waveform (PW) for a number of equal sub-intervals within the predetermined interval;

said signal processor computing a PW gain for generating a normalized PW for each sub-interval and a PW gain vector for the predetermined interval;

a low-pass filter for separating the normalized PW into a slowly evolving waveform (SEW) component and a rapidly evolving waveform (REW) component along every pitch harmonic track;

a voicing measure for characterizing the degree of signal periodicity over the predetermined interval, derived from the input signal, PW, SEW and REW characteristics; and

a quantizer for quantization of the variable dimension SEW magnitude vector by a hierarchical approach, comprising fixed dimension vector quantizers (VQs) for each predetermined interval.

2. A system as recited in claim 1, wherein said quantizer uses a mean-rms-shape decomposition and employs fixed dimension vector quantizers.

3. A system as recited in claim 2, wherein said vector quantizers employ switched backward prediction of the variable dimension SEW magnitude vector components.

4. A system as recited in claim 3, wherein said quantizer provides a switched predictive quantization based on repeating the quantization in both predictive and non-predictive modes and choosing the one that results in minimum quantization distortion.

5. A system as recited in claim 4, wherein said quantizer performs a fixed dimension vector quantization of a SEW mean vector, computed by averaging the SEW harmonic magnitudes over a set of sub-bands that span the bandwidth of the input signal.

6. A system as recited in claim 4, wherein said quantizer performs a joint vector quantization of the root-mean-square (rms) and the fixed dimension shape components of the SEW magnitude vector.

7. A method of joint vector quantization of rms and shape components, comprising:

computation of a fixed dimension SEW mean vector by averaging the SEW harmonic magnitudes over a set of sub-bands that span the bandwidth of the input signal;

computation of a variable dimension SEW deviation vector by subtracting the SEW mean from the SEW magnitude vector;

determination of a fixed dimension SEW deviation subvector from the variable dimension SEW deviation vector based on a dynamic frequency selection approach which uses the quantized LP parameter based spectral characteristics of the input signal during the current interval; and

joint quantization of the rms and shape components of the fixed dimension SEW deviation subvector using a frequency weighted distortion measure.

8. A method of fixed dimension quantization of the mean component of the SEW magnitude vector, comprising:

determination of the full dimension SEW deviation vector from the quantized fixed dimension SEW deviation subvector by padding with zero values or the output of the SEW shape predictor;

computation of a full dimension target SEW vector by sub-band computing the difference between the SEW magnitude vector and the reconstructed full dimension SEW deviation vector;

quantization of the fixed dimension SEW mean vector, so that when converted to a full dimension vector by a piecewise-constant construction, a frequency weighted distortion measure is minimized with respect to the target SEW vector; and

selection of a SEW mean codebook from two possible codebooks based on the degree of voicing using the encoded voicing measure.

9. A method as recited in claim 8, comprising error concealment of the SEW magnitude quantization carried out at the decoder by forcing the mode to be predictive, zeroing out the received SEW mean and shape prediction error vectors.

10. A frequency domain interpolative coding system for magnitude modeling and quantization of prototype waveforms, comprising:

a linear prediction (LP) front end responsive to an input signal providing LP parameters which are quantized and encoded over predetermined intervals and used to compute a LP residual signal;

an open loop pitch estimator responsive to said LP residual signal;

a pitch quantizer;

a pitch interpolator;

said open loop pitch estimator, said pitch quantizer, and said pitch interpolator yielding a pitch contour within the predetermined interval;

a signal processor responsive to said LP residual signal and the pitch contour for extracting a prototype waveform (PW) for a number of equal sub-intervals within the predetermined interval;

said signal processor computing a PW gain for generating a normalized PW for each sub-interval and a PW gain vector for the predetermined interval;

a low-pass filter for separating the normalized PW into a slowly evolving waveform (SEW) component and a rapidly evolving waveform (REW) component along every pitch harmonic track;

a voicing measure for characterizing the degree of signal periodicity over the predetermined interval, derived from input signal, PW, SEW and REW characteristics; and

a fixed dimension vector quantizer for quantization of a variable dimension REW magnitude component using a hierarchical approach for each predetermined interval.

11. A system as recited in claim 10, wherein said vector quantizer performs gain-shape quantization of the variable dimension REW magnitude vector for each predetermined interval, with normalization of the variable dimension REW magnitude vector and separation of the REW magnitude gain and the variable dimension REW magnitude shape vectors.

12. A system as recited in claim 11, wherein quantization of the REW magnitude gain uses the quantized SEW mean vector to compute an estimate of the REW magnitude gain vector and quantizes the estimation error vector using a fixed dimension VQ.

13. A system as recited in claim 11, wherein for each sub-interval, the variable dimension REW magnitude shape vector is transformed into a fixed dimension vector by averaging the normalized REW harmonic amplitudes within a set of sub-bands that span the bandwidth of the input signal.

14. A system as recited in claim 11, wherein the sub-band REW shape vectors are averaged across the predetermined interval to generate an average sub-band REW shape vector for each predetermined interval.

15. A system as recited in claim 10, wherein said vector quantizer provides an error concealment procedure of the REW magnitude by estimating the REW magnitude gain and discarding the received information about the quantized estimation error vector.

16. A frequency domain interpolative coding method for magnitude modeling and quantization of prototype waveforms, comprising:

a linear prediction (LP) front end responsive to an input signal providing LP parameters which are quantized and encoded over predetermined intervals and used to compute a LP residual signal;

an open loop pitch estimator responsive to said LP residual signal;

a pitch quantizer;

a pitch interpolator;

said open loop pitch estimator, said pitch quantizer, and said pitch interpolator yielding a pitch contour within the predetermined interval;

a signal processor responsive to said LP residual signal and the pitch contour for extracting a prototype waveform (PW) for a number of equal sub-intervals within the predetermined interval;

said signal processor computing a PW gain for generating a normalized PW for each sub-interval and a PW gain vector for the predetermined interval;

a low-pass filter for separating the normalized PW into a slowly evolving waveform (SEW) component and a rapidly evolving waveform (REW) component along every pitch harmonic track;

a voicing measure for characterizing the degree of signal periodicity over the predetermined interval, derived from the input signal, PW, SEW and REW characteristics; and

quantization of the variable dimension SEW magnitude vector by a hierarchical approach, using fixed dimension vector quantizers (VQs) for each predetermined interval.

17. A method as recited in claim 16, wherein the quantizing step converts the REW vector to a magnitude-phase form, and encodes a normalized REW magnitude to provide a REW magnitude shape vector comprising multiple sub-band vectors.

18. A method as recited in claim 16, wherein the quantizing step averages the multiple sub-band vectors across time to provide a single average REW sub-band vector for each predetermined period.

19. A method as recited in claim 18, comprising a decoding step for decoding a full-dimension REW magnitude shape vector from the quantized REW sub-band vector employing a piecewise-constant construction.