摘要
针对非限定文本的自动讲话人辨认问题，对提取有效的识别特征进行了理论和实验分析，使用语音信号ＬＰＣ倒谱的瞬时谱和过渡谱的特征参量作为识别特征，运用矢量量化技术，设计并建立了１９人组成的ＡＳＩ系统，为每人分别建立了“动态”和“静态”倒谱系数的两种码本，在连续语音完全不限文本的情况下进行系统的识别实验，得到了较理想的结果。实验表明，上述两种特征矢量均为非常有效的识别特征且可互补结合运用于识别，使性能进一步提高。本文同时还对简单信道变化情况予以初步考察，得到一些有意义的结果。
On the issue of text-independent speaker identification,the extracting effective feature is analysed theoretically and experimentally. The instantaneous and transitionalspectral parameters of LPC cepstrum are uesd as recognition feature. Utilizing vectorquantization technique .an ASI system consisting of 19 speakers is designed and established. For each individual speaker, two cepstral coefficient codebooks, 'static 'and 'dynamic',are constructed respectively. Under the circumstances of continuous and unconstrained utterances,a relatively ideal recognition performance is achieved. The experimental results show that these two features are both very efficient and can be combinedto improve the recognition performance. Simple transmission channel variations are preliminary inspected. Some significient results are obtained.
出处
《山东大学学报（自然科学版）》
CSCD
1994年第2期173-179,共7页
Journal of Shandong University(Natural Science Edition)
关键词
文本无关
矢量量化
语音识别
text-independent
vector quantization
dynamic cepstral coefficient
speechrecognition
