I am a master student and in my final stage of my project. I am using Sphinx 3.5 as the speech recognition engine for developing GUI sw to convert live uttered speech into text and real time video American SIgn Language. From CMU's open source resources i got the CMU dictionary and the trigram dump language model. I've implemented the Sphinx 3.5 into my GUI's sw. I am using Visual Basic for this purpose. Every thing sounds ok, however, the speech recognition accuracy is quite poor, it sounds even less than 50%. An other very weired thing is that sometimes especially when i use the SW at night in live decode mode the accuracy gets better! I have even bought some mike with Active Noice Cancellation and USB pod (it is called Andrea ANC 750), however, the problem of poor accuracy still there. I have went through that parameter optimization as described in Sphinx documentation but without avail. My focus the most is on LiveDecode and LivePretend. Is it about the Acoustic Model, since i am still using the Hub 4 AM that was originally distributed along with the Sphinx3.5 package? Do i have to train my own AM in order to get the best accuracy? Please kindly help me to overcome this very problem?
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I am a master student and in my final stage of my project. I am using Sphinx 3.5 as the speech recognition engine for developing GUI sw to convert live uttered speech into text and real time video American SIgn Language. From CMU's open source resources i got the CMU dictionary and the trigram dump language model. I've implemented the Sphinx 3.5 into my GUI's sw. I am using Visual Basic for this purpose. Every thing sounds ok, however, the speech recognition accuracy is quite poor, it sounds even less than 50%. An other very weired thing is that sometimes especially when i use the SW at night in live decode mode the accuracy gets better! I have even bought some mike with Active Noice Cancellation and USB pod (it is called Andrea ANC 750), however, the problem of poor accuracy still there. I have went through that parameter optimization as described in Sphinx documentation but without avail. My focus the most is on LiveDecode and LivePretend. Is it about the Acoustic Model, since i am still using the Hub 4 AM that was originally distributed along with the Sphinx3.5 package? Do i have to train my own AM in order to get the best accuracy? Please kindly help me to overcome this very problem?