Speaker Feature and Long Window Analysis

次へ: Dependency of classification rate 上へ: Experimental results 戻る: Experimental results

Speaker Feature and Long Window Analysis

It is known from a study of speaker identification that the long-term average spectrum is useful. Therefore, it is expected that the speaker classification rate will be improved by increasing the LPC analysis window length. From such a viewpoint, we examined the classification rate with various LPC analysis window lengths. The universal code book size was set at 64 and 256. In both cases, the frame length ( window interval )was set at half the LPC analysis window length. For each speech data, 16 kinds of random initial models were constructed using random variables. The average for 8 sets of speech data, i.e., the average for 128 trials in total, was taken as the average speaker classification rate. The end condition of HMM training was set to 160 iterations. Other experimental conditions were the same as in experiment 3. Fig.5 shows the results of this experiment.

**図 5:** Relationship between window length and speaker classification rate
$\begin{figure}\begin{center} \fbox{\epsfig{file=PS/figure2.epsf,height=50mm}}%% ,width=70mm}} \end{center}\end{figure}$

The following observations can be obtained from this figure.

When the LPC analysis window length is long, the average classification rate is increases, but decreases beyond a certain point.
The classification rate is the highest when the code book size is 64 and the analysis window length is 341ms.

次へ: Dependency of classification rate 上へ: Experimental results 戻る: Experimental results

Jin'ichi Murakami 平成13年1月19日