Force-aligned transcripts

[Rafael Oliveira]

I'm trying to train a acoustic model with force-aligned transcripts but, I'm
confused about the steps should I take to do it. In the sphinx_train.cfg there
are some properties that refers to this task but, the descriptions provided in
this file were not sufficient to make me understand what should i do. For
example, i can not realize the difference between the properties
falign_ci_mgau and ci_mgau. In short, could you suggest some tutorial,
documentation or information to take me off from the darkness?

[Nickolay V. Shmyrev]

falign_ci_mgau enables multiple gaussian models on forced alignment stage 10
and 11. ci_mgau enables multiple gaussian models on stage 20. That is needed
if you want to tran CI multiple gaussian models for small vocabulary task.

To understand what is gaussian and what are multiple gaussians you can check
any book on HMM-based speech recognition.

[Rafael Oliveira]

Thanks, I read about what you suggested me and now I am trying to train a
model with forced aligned transcripts but I've got a error on the
03.force_align step.

This is the last INFO given in the log:

INFO: main_align.c(919): codDef-art001a: 2360 input frames
lt-sphinx3_align: s3_align.c:927: align_build_sent_hmm: Assertion `stail.predlist' failed.

but, it shows a error before:

ERROR: "main_align.c", line 851: Uttid mismatch: ctlfile = "codDef-art001a"; transcript = "CodDefConsumidor16k/codDef-art001a"

The codDef-art001a is the first transcription in my train.fileids file, and
that is not the first time I'm using it in a train, so to the others programs
that use it there is not any problem with it. Well, the fact is: I do not know
what is happening.

Here is entire log:

INFO:   Initialization of the log add table
INFO:   Log-Add table size = 29356 x 2 >> 0
INFO:   
INFO: feat.c(684): Initializing feature stream to type: '1s_c_d_dd', ceplen=13, CMN='current', VARNORM='no', AGC='none'
INFO: cmn.c(142): mean[0]= 12.00, mean[1..12]= 0.0
INFO:   Reading HMM in Sphinx 3 Model format
INFO:   Model Definition File: /home/20102000030/sphinx/acoustic_model/acoustic_model/model_architecture/db.falign_ci.mdef
INFO:   Mean File: /home/20102000030/sphinx/acoustic_model/acoustic_model/model_parameters/db.falign_ci_cont/means
INFO:   Variance File: /home/20102000030/sphinx/acoustic_model/acoustic_model/model_parameters/db.falign_ci_cont/variances
INFO:   Mixture Weight File: /home/20102000030/sphinx/acoustic_model/acoustic_model/model_parameters/db.falign_ci_cont/mixture_weights
INFO:   Transition Matrices File: /home/20102000030/sphinx/acoustic_model/acoustic_model/model_parameters/db.falign_ci_cont/transition_matrices
INFO: mdef.c(683): Reading model definition: /home/20102000030/sphinx/acoustic_model/acoustic_model/model_architecture/db.falign_ci.mdef
INFO:   Initialization of mdef_t, report:
INFO:   38 CI-phone, 0 CD-phone, 3 emitstate/phone, 114 CI-sen, 114 Sen, 38 Sen-Seq
INFO:   
INFO: kbcore.c(299): Using optimized GMM computation for Continuous HMM, -topn will be ignored
INFO: cont_mgau.c(164): Reading mixture gaussian file '/home/20102000030/sphinx/acoustic_model/acoustic_model/model_parameters/db.falign_ci_cont/means'
INFO: cont_mgau.c(423): 114 mixture Gaussians, 1 components, 1 streams, veclen 39
INFO: cont_mgau.c(164): Reading mixture gaussian file '/home/20102000030/sphinx/acoustic_model/acoustic_model/model_parameters/db.falign_ci_cont/variances'
INFO: cont_mgau.c(423): 114 mixture Gaussians, 1 components, 1 streams, veclen 39
INFO: cont_mgau.c(524): Reading mixture weights file '/home/20102000030/sphinx/acoustic_model/acoustic_model/model_parameters/db.falign_ci_cont/mixture_weights'
WARNING: "cont_mgau.c", line 667: Weight normalization failed for 3 senones
INFO: cont_mgau.c(679): Read 114 x 1 mixture weights
INFO: cont_mgau.c(707): Removing uninitialized Gaussian densities
 0 1 2
WARNING: "cont_mgau.c", line 781: 3 densities removed (3 mixtures removed entirely)
INFO: cont_mgau.c(797): Applying variance floor
INFO: cont_mgau.c(815): 0 variance values floored
INFO: cont_mgau.c(863): Precomputing Mahalanobis distance invariants
INFO: tmat.c(119): Reading HMM transition probability matrices: /home/20102000030/sphinx/acoustic_model/acoustic_model/model_parameters/db.falign_ci_cont/transition_matrices
WARNING: "tmat.c", line 192: Normalization failed for tmat 0 from state 0
WARNING: "tmat.c", line 192: Normalization failed for tmat 0 from state 1
WARNING: "tmat.c", line 192: Normalization failed for tmat 0 from state 2
INFO:   Initialization of tmat_t, report:
INFO:   Read 38 transition matrices of size 3x4
INFO:   
INFO: dict.c(385): Reading main dictionary: /home/20102000030/sphinx/acoustic_model/acoustic_model/falignout/db.falign.dict
INFO: dict.c(388): 12874 words read
INFO: dict.c(393): Reading filler dictionary: /home/20102000030/sphinx/acoustic_model/acoustic_model/falignout/db.falign.fdict
INFO: dict.c(396): 3 words read
INFO: dict.c(429): Added 0 fillers from mdef file
INFO: s3_align.c(1357): logs3(beam)= -460586
ERROR: "main_align.c", line 851: Uttid mismatch: ctlfile = "codDef-art001a"; transcript = "CodDefConsumidor16k/codDef-art001a"
INFO: feat.c(1176): At directory /home/20102000030/sphinx/acoustic_model/acoustic_model/feat
INFO: feat.c(993): Reading mfc file: '/home/20102000030/sphinx/acoustic_model/acoustic_model/feat/CodDefConsumidor16k/codDef-art001a.mfc'[0..-1]
INFO: cmn.c(175): CMN: 11.56 -0.08  0.05  0.25 -0.17 -0.08 -0.11 -0.31 -0.18 -0.29 -0.12 -0.19 -0.18 
INFO: main_align.c(919): codDef-art001a: 2360 input frames
lt-sphinx3_align: s3_align.c:927: align_build_sent_hmm: Assertion `stail.predlist' failed.

Thanks

[Nickolay V. Shmyrev]

The aligner fails to build HMM sequence for your transcription.

Aligner might be not that robust to newlines, spaces and other characters. For
example to UTF-8 BOM symbols and so on. Maybe you want to check the
transcription file and the dictionary to find out if there are any issues like
that.

[Rafael Oliveira]

Thanks for the replay, I will check my transcriptions.