Speech coder and speech decoder

1. A vector quantization apparatus for performing coding of a target vector by multi-stage vector quantization, the apparatus comprising:

a predictor for generating a predictive error vector based on the target vector;

a first codebook for storing a plurality of first code vectors;

a first quantizer for performing a first distance calculation using the predictive error vector provided by the predictor, a first code vector of the plurality of first code vectors stored in the first codebook and the target vector, and performing a first stage of the coding of the target vector using a result of the first distance calculation;

an amplifier storage for storing a plurality of scalars associated with codes of the plurality of first code vectors, respectively, each of the plurality of scalars being obtained by pre-training wherein the pre-training encodes a plurality of speech data by using one of the first code vectors associated with each of the codes and minimizes a sum of encoded distortions of the encoded plurality of speech data; and

a second quantizer for determining a third code vector by multiplying a second code vector stored in a second codebook and one of the plurality of scalars associated with a code of the first code vector determined at the first stage of coding together, calculating a difference vector between the target vector and the first code vector, performing a second distance calculation using the predictive error vector provided by the predictor, the difference vector, and the third code vector, and performing a second stage of the coding of the target vector using a result of the second distance calculation.

2. The vector quantization apparatus according to claim 1 ,

wherein the second code vector stored in the second codebook is obtained by a computation using an amount of sample vectors for learning.

3. A vector quantization method for performing coding of a target vector by multi-stage vector quantization, the vector quantization method being performed with a vector quantizer, the method comprising:

generating, by a predictor, a predictive error vector based on the target vector;

storing a plurality of first code vectors in a first codebook;

storing, in an amplifier storage, a plurality of scalars associated with codes of the plurality of first code vectors, respectively, each of the plurality of scalars being obtained by pre-training wherein the pre-training encodes a plurality of speech data by using one of the first code vectors associated with each of the codes and minimizes a sum of encoded distortions of the encoded plurality of speech data;

reading a first code vector of the plurality of first code vectors from the first codebook, performing a first distance calculation using the predictive error vector provided by the predictor, the first code vector, and the target vector, and performing a first stage of the coding of the target vector using a result of the first distance calculation;

reading one of the plurality of scalars associated with a code of the first code vector determined at the first stage of coding from the amplifier storage;

reading a second code vector from a second codebook, and determining a third code vector by multiplying the second code vector and the read one of the plurality of scalars together;

calculating a difference vector between the target vector and the first code vector;

performing a second distance calculation using the predictive error vector provided by the predictor, the difference vector, and the third code vector, and performing a second stage of the coding of the target vector using a result of the second distance calculation.

4. The vector quantization method according to claim 3 ,

wherein the second code vector stored in the second codebook is obtained by a computation using an amount of sample vectors for learning.