## 4aSC39. Quantization of vector sequences using statistics of neighboring input vectors.

### Session: Thursday Morning, December 5

### Time:

**Author: Keiichi Tokuda**

**Location: Dept. of Intelligence and Comput. Sci., Nagoya Inst. of Technol., Gokiso-cho, Showa-ku, Nagoya, 466 Japan**

**Author: Takao Kobayashi**

**Location: Tokyo Inst. of Technol., Yokohama, 226 Japan**

**Author: Takashi Masuko**

**Location: Tokyo Inst. of Technol., Yokohama, 226 Japan**

**Author: Satoshi Imai**

**Location: Tokyo Inst. of Technol., Yokohama, 226 Japan**

**Abstract:**

A vector quantization method using statistics, i.e., mean and covariance,
of neighboring input vectors (or linear transform of those) is proposed. In the
proposed method, the distance between the input vector and the codeword is
measured by the output probability defined by the statistics, and the output
vector sequence is determined so that the output probability of the output
vector sequence is maximized. The codebook can be trained by a conventional
training procedure based on the statistically defined distance or a simplified
distance. The value of the output vector sequence is determined by a high-order
set of equations. Fortunately, it can be shown that the set of equations can be
solved by a fast algorithm. In the conventional vector quantization methods,
adjoining output vectors can change discontinuously; quantizing speech spectral
parameter vectors causes perceivable glitches in the synthesized speech,
whereas, in the proposed method, the change of adjoining output vectors is
controlled appropriately according to the statistics of neighboring input
vectors. Through an example of vector quantization of the LSP coefficients
obtained from natural speech, it is shown that the proposed method can improve
objective and subjective performance of vector quantization.

ASA 132nd meeting - Hawaii, December 1996