Method for improving speech quality in code excited linear predictive speech coding

• US 4,980,916 A
• Filed: 10/26/1989
• Issued: 12/25/1990
• Est. Priority Date: 10/26/1989
• Status: Expired due to Term

• Current Assignee: Lockheed Martin Corporation (Martin Marietta Corporation)
• Original Assignee: General Electric Company
• Inventors: Zinser, Richard L.
• Primary Examiner(s): KEMENY, EMANUEL
• Application Number:
  US07/427,074
• Time in Patent Office: 425 Days
• Field of Search: 381/36, 381/31, 381/38, 381/47, 381/49
• US Class Current: 704/207
• CPC Class Codes: G10L 19/12 the excitation function being a code excitation, e.g. in code excited linear prediction [CELP] vocoders
• CPC Class Codes: G10L 2019/0003 Backward prediction of gain
• CPC Class Codes: G10L 2019/0011 Long term prediction filters, i.e. pitch estimation
• CPC Class Codes: G10L 2019/0013 Codebook search algorithms
• CPC Class Codes: G10L 25/06 the extracted parameters being correlation coefficients

Abstract

By reconciling differences between the estimator and the filter of a code excited linear predictive (CELP) voice coder, higher quality is achieved in the output speech. The pulse amplitudes and pitch tap gain are solved for simultaneously to minimize the estimator bias in the CELP excitation. Increased signal to noise ratio is accomplished by modifying the pitch predictor such that the pitch synthesis filter accurately reflects the estimation procedure used to find the pitch tap gain, and by improving the excitation analysis technique such that the pitch predictor tap gain and codeword gain are solved for simultaneously, rather than sequentially. These modifications do not result in an increased transmission rate or significant increase in complexity of the CELP coding algorithm.

Claims

1. A method for improving speech quality in code excited linear predictive voice coders, comprising the steps of:

determining a pitch predictor tap gain as a normalized cross-correlation of an input sequence and pitch buffer samples by copying previous samples at a distance of P samples so as to extend pitch buffer length;

modifying a pitch synthesis filter so that a pitch predictor output sequence is a series computed for each interval P; and

simultaneously solving for pulse amplitudes and pitch tap gain, thereby minimizing estimator bias in the code excitation.

3. The code excited linear predictive coder recited in claim 1 further comprising linear predictive synthesis filter means responsive to said linear predictive coefficients and said combined excitation for generating an output signal that closely resembles said input signal.

2. A code excited linear predictive coder comprising:

linear predictive code analysis means for receiving an input signal and generating from said input signal a set of linear predictive filter coefficients;

weighting means for receiving said input sequence and said set of linear predictive filter coefficients for generating a weighted input sequence;

codebook means for generating output codewords;

first weighted linear predictive synthesis filter means responsive to said set of linear predictive filter coefficients and said codewords for generating synthesis filtered codewords;

pitch filter means for generating pitch excitation sequences;

second weighted linear predictive synthesis filter means responsive to said set of linear predictive filter coefficients and said pitch excitation sequences for generating synthesis filtered pitch excitation sequences;

equation solving means receiving said weighted input sequence, said synthesis filtered codewords and said synthesis filtered pitch excitation sequences for computing a pitch predictor tap gain and a codeword excitation gain;

first multiplying means for multiplying said codebook output sequences by said codeword excitation gain to produce a codebook excitation output signal;

second multiplying means for multiplying said pitch excitation sequences by said pitch predictor tap gain to produce a pitch predictive excitation; and

summing means for summing said codebook excitation output signal and said pitch predictive excitation to generate a combined excitation to be transmitted with said linear prediction coefficients.

4. A method of generating an excitation sequence for transmission with linear predictive coefficients of an input signal in a code excited linear predictive speech coder, comprising the steps of:

computing a pitch lag by finding the location of a maximum cross-correlation between a weighted input sequence and synthesis-filtered contents of a pitch buffer of the coder;

generating an unscaled pitch prediction sequence using the computed pitch lag and a pitch tap gain of 1.0;

passing the unscaled pitch prediction sequence through a weighted linear predictive synthesis filter to produce an unscaled weighted synthesis pitch prediction sequence;

computing a pitch prediction sequence variance from the unscaled weighted synthesis pitch prediction sequence and a cross-correlation between the weighted input sequence and unscaled weighted synthesis pitch prediction sequence;

conducting an exhaustive Gaussian codebook search and, for each codeword output sequence obtained from said codebook, computing a codeword output sequence variance and a cross-correlation between the codeword output sequence and the weighted input sequence;

determining optimal values for codeword gain and pitch tap gain from said computed variances and said cross-correlations;

multiplying the pitch prediction sequence by the optimal value of pitch tap gain to arrive at a scaled pitch prediction sequence;

multiplying the codeword output sequence by the optimal codeword gain to arrive at a scaled codeword sequence; and

summing the scaled pitch and codeword sequences to generate parameters representing said excitation sequence.

5. The method of generating an excitation sequence as recited in claim 4 further comprising the step of transmitting said parameters representing an excitation sequence together with said linear prediction coefficients.

6. The method of generating an excitation sequence as recited in claim 5 further comprising the step of utilizing said excitation sequence and said linear prediction coefficients for synthesizing an output signal which closely resembles said input signal.