WIP: Exponential smoothing CMN

[vimalmanohar]

Adding exponential smoothing based CMN where the moving average is estimated using a exponential smoothing with a time decay and a latency at the beginning of the utterance.

[RuABraun]

Is there a reason this seems to have been abandoned and the focus is instead on the normal rectangular online cmvn?

[danpovey]

That's a good question. @vimalmanohar I think I asked you before about this but didn't get a response. Can you please let us know if there is any performance difference?
The exponential one is easier to implement so might be preferable if performance is the same.

[vimalmanohar]

@phanisankar-nidadavolu was also testing this. I think exponential CMN is about the same as sliding CMN, but the performance degrades when used with i-vectors (ivectors alone is better than ivectors + CMN).

[danpovey]

So you're saying that with i-vectors, exponential is worse than sliding-window CMN? It seems to me that the right way to do this is to apply the same type of CMN when extracting the ivectors, and to do this for the features used to estimate the ivectors, not just the ones used to calculate the posteriors as we normally do. Did you do this?

…

On Sat, Aug 24, 2019 at 10:03 AM Vimal Manohar ***@***.***> wrote: @phanisankar-nidadavolu <https://github.com/phanisankar-nidadavolu> was also testing this. I think exponential CMN is about the same as sliding CMN, but the performance degrades when used with i-vectors (ivectors alone is better than ivectors + CMN). — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#3157?email_source=notifications&email_token=AAZFLO374FTGDBMFFPV6BLTQGFSUZA5CNFSM4HAIYUSKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD5CD35Q#issuecomment-524566006>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAZFLO5P6N2PYL2YMGRQYKTQGFSUZANCNFSM4HAIYUSA> .

[vimalmanohar]

No. With i-vectors, exponential is worse than no-CMN. CMN is not applied when estimating ivectors. The ivectors are still extracted in the conventional way. I think Phani was trying to use ivectors extracted with CMN features on augmented and non-augmented data. I'm not sure if he got any results on this. On Sat, Aug 24, 2019 at 1:17 PM Daniel Povey <[email protected]> wrote:

…

So you're saying that with i-vectors, exponential is worse than sliding-window CMN? It seems to me that the right way to do this is to apply the same type of CMN when extracting the ivectors, and to do this for the features used to estimate the ivectors, not just the ones used to calculate the posteriors as we normally do. Did you do this? On Sat, Aug 24, 2019 at 10:03 AM Vimal Manohar ***@***.***> wrote: > @phanisankar-nidadavolu <https://github.com/phanisankar-nidadavolu> was > also testing this. I think exponential CMN is about the same as sliding > CMN, but the performance degrades when used with i-vectors (ivectors alone > is better than ivectors + CMN). > > — > You are receiving this because you commented. > Reply to this email directly, view it on GitHub > < #3157?email_source=notifications&email_token=AAZFLO374FTGDBMFFPV6BLTQGFSUZA5CNFSM4HAIYUSKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD5CD35Q#issuecomment-524566006 >, > or mute the thread > < https://github.com/notifications/unsubscribe-auth/AAZFLO5P6N2PYL2YMGRQYKTQGFSUZANCNFSM4HAIYUSA > > . > — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#3157?email_source=notifications&email_token=ABABGV2YHJIOWQWOGOI5U4LQGFULDA5CNFSM4HAIYUSKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD5CEDGQ#issuecomment-524566938>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABABGV4WY3SB6F3M2TH6GMTQGFULDANCNFSM4HAIYUSA> .

-- Vimal Manohar PhD Student Center for Language and Speech Processing Johns Hopkins University Baltimore, MD

[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.