search

k2-fsa / snowfall

Moved to https://github.com/k2-fsa/icefall
Apache License 2.0
144 stars 42 forks source link

Topology problem? #200

Closed danpovey closed 3 years ago

danpovey commented 3 years ago

There seems to be an issue with the CTC topology, it's not quite as it should be

labels= <eps> SIL SIL SIL SIL SIL SIL SIL SIL SIL SIL F ER0 ER0 ER0 SIL AE1 AE1 T <eps> DH DH AH0 SIL <eps> P P EH1 EH1 EH1 R AH0 AH0 L L SIL SIL AH1 V V HH HH ER0 ER0 SIL SIL OW1 OW1 N N <eps> IH0 IH0 G G <eps> Z Z IH1 S S T T AH0 N N S S SIL SIL SIL SIL SIL SIL SIL SIL SIL AH0 N D <eps> W W EH1 N DH DH IY0 SIL AH1 AH1 T T ER0 <eps> ER0 ER0 HH HH AE1 AE1 L L IH0 IH0 N N Z Z HH HH AE1 D <eps> SIL SIL SIL D D IH0 Z Z ER1 ER1 ER1 T IH0 IH0 D D SIL HH HH ER0 SIL SIL SIL SIL SIL SIL SIL SIL SIL SIL SIL SIL <eps> SH SH SH IY1 IY1 IY1 SIL <eps> R IH0 IH0 P P P EH1 EH1 L <eps> L D D DH AH0 IH0 N <eps> <eps> V EY1 EY1 D D ER0 <eps> ER0 <eps> SIL SIL SIL SIL SIL SIL SIL SIL SIL SIL SIL AH0 AH0 N D SIL AH1 V <eps> <eps> HH HH ER0 ER0 SIL <eps> OW1 OW1 OW1 N N <eps> AH0 AH0 K K AO1 AO1 AO1 R R D D SIL SIL G G EY1 EY1 EY1 V V V <eps> SIL L L IH1 B ER0 ER0 ER0 T T IY0 IY0 IY0 <eps> T T UW1 UW1 SIL SIL AO1 AO1 AO1 L L SIL DH DH AH0 SIL N N EY1 EY1 SH AH0 N N N Z Z W IH0 TH TH IH1 IH1 N N <eps> <eps> DH AH0 <eps> P P IH1 IH1 L L ER0 ER0 ER0 Z <eps> Z Z SIL SIL SIL SIL SIL SIL SIL SIL SIL SIL SIL
aux_labels= FOR AT THE PERIL OF HER OWN EXISTENCE AND WHEN THE UTTER HALLIN'S HAD DESERTED HER SHE REPELLED THE INVADER AND OF HER OWN ACCORD GAVE LIBERTY TO ALL THE NATIONS WITHIN THE PILLARS

See above, ER0 <eps> ER0 and Z <eps> Z Z. In the best-path FSA from decoding, there are instances where a phone has an in the middle and then it's repeated. I believe that is not allowed unless the phone itself is repeated, which it doesn't seem to be here, I checked the pron of PILLARS for instance, there is just one Z.

csukuangfj commented 3 years ago

@danpovey

Could you attach the resulting best path as a .pt file? I would like to visualize it.

danpovey commented 3 years ago

sorry don't have it as .pt... I should have mentioned, this was with LM rescoring, it's possible that the epsilons come from there. In fact I am running without LM rescoring and I don't see such issues. So that may be the issue.