[src] Modified mel computations (cleaned up version)

[danpovey]

to be tested. Would be activated by setting
modified=true
in conf/mfcc_hires.conf

[RuABraun]

This is supposed to give a small but consistent performance improvement correct?

[danpovey]

Yes. I'd say around 0.2 to 0.3% absolute on average (at least when the WER is in the 20 to 30% range). It needs some more testing before I'm ready to merge it. Of course if you want to help, that's great.

…

On Mon, Mar 25, 2019 at 9:00 AM Rudolf A. Braun ***@***.***> wrote: This is supposed to give a small but consistent performance improvement correct? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#3165 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADJVu0X4FSGLo407ljIiEwtllhONQoXgks5vaMiLgaJpZM4cFx2K> .

[danpovey]

And by testing, I am talking about the effect on WER, not just does it run.

…

On Mon, Mar 25, 2019 at 10:40 AM Daniel Povey ***@***.***> wrote: Yes. I'd say around 0.2 to 0.3% absolute on average (at least when the WER is in the 20 to 30% range). It needs some more testing before I'm ready to merge it. Of course if you want to help, that's great. On Mon, Mar 25, 2019 at 9:00 AM Rudolf A. Braun ***@***.***> wrote: > This is supposed to give a small but consistent performance improvement > correct? > > — > You are receiving this because you authored the thread. > Reply to this email directly, view it on GitHub > <#3165 (comment)>, > or mute the thread > <https://github.com/notifications/unsubscribe-auth/ADJVu0X4FSGLo407ljIiEwtllhONQoXgks5vaMiLgaJpZM4cFx2K> > . >

[RuABraun]

I could do a build on my own data (so you'd have to take my word for the results), if that works for you?

[danpovey]

Yes, that works for me.

…

On Tue, Mar 26, 2019 at 8:51 AM Rudolf A. Braun ***@***.***> wrote: I could do a build on my own data (so you'd have to take my word for the results), if that works for you? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#3165 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADJVuzTcXb0l21aXMGdn85kBTN76gxqIks5vahfjgaJpZM4cFx2K> .

[pegahgh]

It should be breakpoint_ instead of 700.

[danpovey]

I was thinking of making this a separate issue from the breakpoint, so that when we tune it it doesn't have to have the same value. Can you compare 700 with 300 (our current breakpoitn) in your new experiments? If 300 works better we can reconsider.

[RuABraun]

Sorry about the delay on this:

Some info about the model: 100h with 13 layer tdnnf. No ivector. Decoding with a 1M ngram language model (no rescoring).

Because of how my data is acquired, I cannot always guarantee that there is no overlap in speakers between training/test, these are the test sets A, B, C. Some test sets I can guarantee it, those are X, Y, Z

	X	Y	Z	A	B	C
baseline	18.01	17.57	29.12	19.57	12.03	23.82
modified	18.70	17.85	29.56	19.48	12.02	23.59

As can see from the results it seems the modified version overfits more.

[danpovey]

Great, thanks. It turns out there may have been a problem with that particular code (i.e. there was a mismatch to what it should have been), although I don't know how important that actually was.
Hopefully at some point in the future I will have a fixed version.

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[stale]

This issue has been automatically closed by a bot strictly because of inactivity. This does not mean that we think that this issue is not important! If you believe it has been closed hastily, add a comment to the issue and mention @kkm000, and I'll gladly reopen it.

[stale]

This issue has been automatically marked as stale by a bot solely because it has not had recent activity. Please add any comment (simply 'ping' is enough) to prevent the issue from being closed for 60 more days if you believe it should be kept open.