Hi
I tried the same 26 words now with the phone model . But the problem which it gave me was at the end when i used the sphinx decoder---Illegal
memory access error .
There was also a core dump file of 2.9 MB which says BW might have crashed. The scripts as usual only showed the same warnings & errors in the first iteration-
WARNING: "gauden.c", line 1382: (mgau= 0, feat= 2, density= 249) never observed
WARNING: "gauden.c", line 1382: (mgau= 0, feat= 2, density= 253) never observed
WARNING: "gauden.c", line 1382: (mgau= 0, feat= 2, density= 254) never observed
INFO: main.c(695): Normalizing var
ERROR: "gauden.c", line 1424: var (mgau= 0, feat= 2, density=127, component=0) < 0
ERROR: "gauden.c", line 1424: var (mgau= 0, feat= 2, density=127, component=1) < 0
ERROR: "gauden.c", line 1424: var (mgau= 0, feat= 2, density=127, component=2) < 0
ERROR: "gauden.c", line 1424: var (mgau= 0, feat= 2, density=220, component=1) < 0
ERROR: "gauden.c", line 1424: var (mgau= 0, feat= 2, density=220, component=2) < 0
INFO: s3mixw_io.c(238): Wrote
/Speech/speech/SphinxTrain/robot/model_parameters/robot.ci_semi/mixture_weights [170x4x256 array]
INFO: s3tmat_io.c(180): Wrote
/Speech/speech/SphinxTrain/robot/model_parameters/robot.ci_semi/transition_matrices [34x5x6 array]
INFO: s3gau_io.c(218): Wrote
/Speech/speech/SphinxTrain/robot/model_parameters/robot.ci_semi/means [1x4x256 array]
INFO: s3gau_io.c(218): Wrote
/Speech/speech/SphinxTrain/robot/model_parameters/robot.ci_semi/variances [1x4x256 array]
Current Overall Likelihood Per Frame = 1.9822893858661
I had only 4 iterations in BW training
What could be problem ?
And Kmeans log showed
Warning: Aborting kmeans; bad initialisation
Is it a problem with my way of recording audio data . What i am doing is asking the speaker to say the word once and save it as a wave file(16 bit
, 16KHz ,PCM format) and then convert it into raw format using Sox . The phrase( No. of 1 word phrases= 14 & No. of 2 word phrases = 12) is
repeated 4 times and saved as 4 files . Similarly i'm using 6 different speakers(3 male and 3 female) and asking them to repeat the same procedure.
Is it right ?
I think the problem lies in the audio training data itself . Any suggestions over this problem ?
Please help me out !
Thanks in advance
Edison
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi
I tried the same 26 words now with the phone model . But the problem which it gave me was at the end when i used the sphinx decoder---Illegal
memory access error .
There was also a core dump file of 2.9 MB which says BW might have crashed. The scripts as usual only showed the same warnings & errors in the first iteration-
WARNING: "gauden.c", line 1382: (mgau= 0, feat= 2, density= 249) never observed
WARNING: "gauden.c", line 1382: (mgau= 0, feat= 2, density= 253) never observed
WARNING: "gauden.c", line 1382: (mgau= 0, feat= 2, density= 254) never observed
INFO: main.c(695): Normalizing var
ERROR: "gauden.c", line 1424: var (mgau= 0, feat= 2, density=127, component=0) < 0
ERROR: "gauden.c", line 1424: var (mgau= 0, feat= 2, density=127, component=1) < 0
ERROR: "gauden.c", line 1424: var (mgau= 0, feat= 2, density=127, component=2) < 0
ERROR: "gauden.c", line 1424: var (mgau= 0, feat= 2, density=220, component=1) < 0
ERROR: "gauden.c", line 1424: var (mgau= 0, feat= 2, density=220, component=2) < 0
INFO: s3mixw_io.c(238): Wrote
/Speech/speech/SphinxTrain/robot/model_parameters/robot.ci_semi/mixture_weights [170x4x256 array]
INFO: s3tmat_io.c(180): Wrote
/Speech/speech/SphinxTrain/robot/model_parameters/robot.ci_semi/transition_matrices [34x5x6 array]
INFO: s3gau_io.c(218): Wrote
/Speech/speech/SphinxTrain/robot/model_parameters/robot.ci_semi/means [1x4x256 array]
INFO: s3gau_io.c(218): Wrote
/Speech/speech/SphinxTrain/robot/model_parameters/robot.ci_semi/variances [1x4x256 array]
Current Overall Likelihood Per Frame = 1.9822893858661
I had only 4 iterations in BW training
What could be problem ?
And Kmeans log showed
Warning: Aborting kmeans; bad initialisation
Is it a problem with my way of recording audio data . What i am doing is asking the speaker to say the word once and save it as a wave file(16 bit
, 16KHz ,PCM format) and then convert it into raw format using Sox . The phrase( No. of 1 word phrases= 14 & No. of 2 word phrases = 12) is
repeated 4 times and saved as 4 files . Similarly i'm using 6 different speakers(3 male and 3 female) and asking them to repeat the same procedure.
Is it right ?
I think the problem lies in the audio training data itself . Any suggestions over this problem ?
Please help me out !
Thanks in advance
Edison