next up previous
次へ: Conclusion 上へ: Spontaneous Speech Recognition 戻る: Experiments for spontaneous speech

Results of spontaneous speech


In the filled-pause skip experiments, the number of filled-pauses in the vocabulary was 109. The experimental results are shown in Table  4 .


sentence recognition rate(%)
表 4: Results of spontaneous speech recognition
speaking order original Filled-pauses Phone-strings  
style   word trigram skip skip  
read 1 90.5% 82.1% 92.7%  
  $\sim$ 8 98.5% 95.4% 98.9%  
semi 1 42.0% 63.7% 65.6%  
spontaneous $\sim$ 8 43.9% 80.1% 72.9%  
spontaneous 1 25.3% 42.7% 42.0%  
  $\sim$ 8 39.5% 61.1% 63.4%  

speaker-independent; word trigram; beam width: 16384; $\alpha $: 16;

As can be seen, the phone-strings skip method obtains a higher performance than the filled-pauses skip method. In addition, the sentence recognition rate for spontaneous speech is 42.0% as opposed to 25.3% with no such procedure; including the semantically correct sentences, the sentence recognition rate was about 75%. This means that this algorithm is effective in spontaneous speech recognition.



next up previous
次へ: Conclusion 上へ: Spontaneous Speech Recognition 戻る: Experiments for spontaneous speech
Jin'ichi Murakami 平成13年1月19日