Accent Issues in Large Vocabulary Continuous Speech Recognition (lvcsr) Chao Huang Eric Chang Tao Chen
Download 282.44 Kb.
|
- Navigate this page:
- Table 2 .1
- Table 2 .2
- Test Sets Accent Speakers
- Table 2.3
- 21.75 28.28 BES(1000BJ+500SH) 8.56
Cross Accent ExperimentsIn order to investigate the impact of accent on the state of the art speech recognition system, we have carried lots of experiments based on Microsoft Chinese speech engine, which has been successfully delivered into Office XP and SAPI. In addition to many kinds of mature technologies such as Cepstrum Mean Normalization, decision tree based state tying, context dependent modeling (triphone) and trigram language modeling, which are all been testified to be important and adopted in the system, tone related information, which are very helpful to be distinguished for Asian tonal language, have also been integrated into out baseline system through including pitch and delta pitch into feature streams and detailed tone modeling. In one word, all improvements and results shown here are achieved based on a solid and powerful baseline system. The details about experiment and results are listed as follows: Experiments setup
Table 2.1: Summary of training corpora for cross accent experiments, Here BJ, SH and GD means Beijing, Shanghai and Guangdong accent respectively.
Table 2.2: Summary of test corpora for cross accent experiments, PPc show here is character perplexity of test corpora according to the LM of 54K.Dic and BG=TG=300,000.
Table 2.3: Character error rate for cross accent experiments.
It is easily concluded from Table 2.3 that accent is a big problem that impacts the state of the art speech recognition systems. Compared with accent specific model, cross accent model may increase error rate by 40-50%.
| |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||