ctc with context dependent phones

[xfwu]

Hi

For ctc with context dependent phones, is there any special thing I should do. Can i just use the acoustic model's output ( CD phones) and sum up the probabilities of these share the same middle phone, then use the TLG script in this project?

Best

[fmetze]

By default, if you want to decode with the CD decoding graph, you would need to create the counts for the CD units in the training data, and use those to scale the posteriors. I think there is some possibility of “smoothing”, maybe, i.e. of averaging the counts of CD phones with the same center phones, but not sure if you want to set them all to the same value. I would be happy to look at your experimental results in that area and discuss, this is interesting, though.

On Aug 28, 2017, at 6:43 AM, Xiaofeng Wu [email protected] wrote:

Hi

For ctc with context dependent phones, is there any special thing I should do. Can i just use the acoustic model's output ( CD phones) and sum up the probabilities of these share the same middle phone, then use the TLG script in this project?

Best

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/srvk/eesen/issues/145, or mute the thread https://github.com/notifications/unsubscribe-auth/AEnA8TuOSifMFE_K5vpJ2NJQckiWqQ1Vks5scpnBgaJpZM4PEXAO.