Hi,i have used the pocketsphinx to build an short isolated commands recognition application(consist of 20 commands,each command has 1-2 words,such as "sing song","light up" ).And the result works well,but there are some question to be solved:
1,I just use the build-in model,its smallest unit is phoneme.If I train the model based on word, will the result be better?
2,After having speak a command,there are little long time to display the result.How can i do to reduce the time of recognition?
3,I read some paper to find the confidence scores can help to reject the oov. In the pocketsphins,i know it's hard to have the confidence scores for continuous speech recognition. Are there some method to get the confidence scores for my short isolated commands recognition?
Thanks in advance.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
1,I just use the build-in model,its smallest unit is phoneme.If I train the model based on word, will the result be better?
Smallest unit is "state", you can learn more in our tutorial http://cmusphinx.sourceforge.net/wiki/tutorialconcepts. It matters which state do you put in HMM. Some toolkits allow you to specify words with large amount of states, this practice is not supported by pocketsphinx, you can simply use words with phones instead, they are equivalent.
2,After having speak a command,there are little long time to display the result.How can i do to reduce the time of recognition?
You can take partial results into account.
3,I read some paper to find the confidence scores can help to reject the oov. In the pocketsphins,i know it's hard to have the confidence scores for continuous speech recognition. Are there some method to get the confidence scores for my short isolated commands recognition?
It is better to avoid random paper and read textbooks instead. Confidence scores from short commands are not easy to get overall. If you are looking for continuous speech recognition, it's better to check our tutorial about keyword spotting.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi,i have used the pocketsphinx to build an short isolated commands recognition application(consist of 20 commands,each command has 1-2 words,such as "sing song","light up" ).And the result works well,but there are some question to be solved:
1,I just use the build-in model,its smallest unit is phoneme.If I train the model based on word, will the result be better?
2,After having speak a command,there are little long time to display the result.How can i do to reduce the time of recognition?
3,I read some paper to find the confidence scores can help to reject the oov. In the pocketsphins,i know it's hard to have the confidence scores for continuous speech recognition. Are there some method to get the confidence scores for my short isolated commands recognition?
Thanks in advance.
Smallest unit is "state", you can learn more in our tutorial http://cmusphinx.sourceforge.net/wiki/tutorialconcepts. It matters which state do you put in HMM. Some toolkits allow you to specify words with large amount of states, this practice is not supported by pocketsphinx, you can simply use words with phones instead, they are equivalent.
You can take partial results into account.
It is better to avoid random paper and read textbooks instead. Confidence scores from short commands are not easy to get overall. If you are looking for continuous speech recognition, it's better to check our tutorial about keyword spotting.