Hello, I'm new in ASR making and I am getting very low accuracy in approximately 12.07% but during training and decoding, I am getting Word Error Rate of 24.7% which is too high also.
I already tried using bestpath and -kws during pocketsphinx_batch and I get the same result. I am trying to use higher (more) numbers of gaussians to make the Word Error Rate lower and try to get Accuracy higher. By the way, I am making an ASR which recognizes digits only.
Thank you in advance.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Thanks a lot for the reply Sir Nickolay, I am not able to share the training and test data. But possibly is there a difference in what is the decoder configuration in training and when using the pocketsphinx_batch?
Can I possible make the senones higher than 200 but lower than 1000 because I'm using cont. type and have just 8 hours of training data? I would be playing the number of senones and it might take a while. Thank you Nickolay for always giving me solutions and ideas to what must be done to improve and resolve some issues. I look forward for the error rate to just become less than 5% and have its accuracy to be 95% or more.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hello, I'm new in ASR making and I am getting very low accuracy in approximately 12.07% but during training and decoding, I am getting Word Error Rate of 24.7% which is too high also.
I already tried using bestpath and -kws during pocketsphinx_batch and I get the same result. I am trying to use higher (more) numbers of gaussians to make the Word Error Rate lower and try to get Accuracy higher. By the way, I am making an ASR which recognizes digits only.
Thank you in advance.
You need to share your training and test data to get help on this issue.
Thanks a lot for the reply Sir Nickolay, I am not able to share the training and test data. But possibly is there a difference in what is the decoder configuration in training and when using the pocketsphinx_batch?
Can I possible make the senones higher than 200 but lower than 1000 because I'm using cont. type and have just 8 hours of training data? I would be playing the number of senones and it might take a while. Thank you Nickolay for always giving me solutions and ideas to what must be done to improve and resolve some issues. I look forward for the error rate to just become less than 5% and have its accuracy to be 95% or more.