Totally wrong prediction during testing on pretrained model

[ChenyuLInx]

Hi there,

Thanks for the great implementation. I tried to use the LibriSpeech pre-trained model you provided on the website. When I test it with my own data. I always got the wrong prediction. Any idea why the performance is so off?

Here's my result:

smssysysysdysmsmsysmsdsdsdysydsssskskpsbsmsbysyssbslbsbbysllssmsbdsgsdskssdsdsdsbsdsysysdsmlsmsdsyskysydsydyybymsbysssmsysmsmsysdbsdsdsyysssysmsbsbysysbmsmsbmbsbsmsysysmsbsmsmsbsbsysdksdlsmrusmsdssypsksdsbsdsdsydsysxspsmsmsmsysmbsysmsbybkdysksdysmssdsysmsssbssnyskskspssmsdypbyssdlskysmsdsysymsmsmspsdsysbsmsmsysmsdydsxsbsmpmssmsysdysysmysmsxsdsyyysbdsksdssysmsmsxsmsdsssmsmsmsmsrssspsdsbssmbsmsssdydsymysbmsmsdysbsbssmsdssddssddsdsdsdsdssspsddysdssspmbspsydsbsispsyysdsysyspsysysysdsysddpdbyssdssbsywssdsdpsbmsdspdsdssssmsmsmsmsbsmssmsmsbsyssdsbsdsdsdkdsddsdydbddksymbsbsdsxspssmspsylsbsmsyscssmsmsdysbsysysydssbsksmsybsmsmsxssmssysmsdmpsmspsdsddsdssdksmsdsbdsdgysgkpsmsmsmsdbsdsdydssdysmsmsbssybssmsmslsmsnxmsmbpsbsdyklsmpsmsdsdsssysyssbsbsbspsdsmsmsmspsdsdydgdsydkdsdsssbsbsbsbsdsbmbbsspsmsmbsmpbsdysyspsbsddyssysuspssysbsbsyspsdsdssmsmsmsybsksysyksxkyssysysydybsdpssdssysmsmsmymsmsbsdsdysysksdsdssdsmsdylslbdybdsykfsmssmsmsmsmsmsysksskskdsskssssdysssdsyskydsdyssdddsdsdsysysbsdsdksdsdslsysysysmsmsmspsmsdsysysmsmsmsmsdysmsmssbsysdysdssusmsmsmsmsmsmsdspslslspsdsbyssybsspbsdssdmsdsmsbsysdsdssmsbsbspsdisbsbssmsbsdsdssymsbsmsmdspsdsssmsbspysbsdysbsyssmsmrumsdpsysysyssysysysdsdyyssssmbspsisbskssmsysydydsdbdbsbbd

Thank you so much. Chenyu

[ArtemiyFirsov]

Yes, same for me ( sssmsymsmsbsbsbsblmsmsbsdbysssdmlsdssrsmsmsmsysmsmspsysmsmsmsmsmsysdsdsydsmsmsmsmsmsmsmsmsmsdsdsmsmsdymsdysmsmsmsmsysmsmsmsmsmsmsmpsmsmsdysysbsdssdssg

Maybe it is underfitting problem? Converted audio to .wav 16kHz with 1 channel an tried 8bit and 16bit. Same everywhere.

[shantanudev]

@SeanNaren Yes, I am also facing the same exact issue since I have had to update my system. Is this a problem with the CTC function?