htk error and solution list
UNDERSTANDING HTK ERROR MESSAGES
Various problems & solutions I've come across in using HTK for building a WSJ recognizer and for my dissertation work in Language Modeling. If you're here to find answers for your own project, consider posting your problems & solutions on your own website, for others to learn from, too. |
PROBLEM | SOLUTION |
HLEd -d prondict -i monophone.mlf mkphones0.led words.mlf Does nothing, only #!MLF!# is returned in the output. |
There need to be double quotes around the lab filename in the words.mlf file: "*/xxx.LAB" instead of '*/xxx.lab' |
HDMan -l hdman.log -w lists/all.wordlist lists/all.words.monophones.dict lists/cmudict.sort ERROR [+1452] ReadDictProns: word A out of order in dict lists/cmudict.sort FATAL ERROR - Terminating program HDMan |
Unix sort doesn't seem to match the sort HTK is looking for. Python's sort function seems to work. Numbers are sorted with '.' before 0, shorter before longer (1 < 1.0 < 10 < 100) |
HLEd -l '*' -d lists/allwords.prons.dict -i lists/all.phonemlf src/mkphones0.led lists/all.wordmlf ERROR [+5013] ReadString: String too long FATAL ERROR - Terminating program HLEd |
Make changes to the pronunciation dictionary: Replace all multiple spaces with single space; Replace all tabs with single space; Put a '\' before every double quote ("); %" Put a '\' before any dictionary entry beginning with single quote (') |
HLEd -l '*' -d lists/allwords.prons.dict.notabnospace -i lists/all.phonemlf src/mkphones0.led lists/all.wordmlf ERROR [+1232] NumParts: Cannot find word ~ in dictionary FATAL ERROR - Terminating program HLEd |
Add that word to the dictionary, resort if necessary |
ERROR [+1232] NumParts: Cannot find word MR. STEINBERG in dictionary FATAL ERROR - Terminating program HLEd |
In the MLF file, the line "MR." ended with a slash, remove the slash from the MLF file. |
HLEd -l '*' -d prondict -i train.monophone.mlf mkphones0.led train.rem.mlf ERROR [+6550] LoadHTKList: Label Name Expected FATAL ERROR - Terminating program HLEd |
For all numbers in train.rem.mlf, precede them with \ so they don't look like a time. |
HLEd -d train.prondict -i train.monophone.mlf mkphones0.led tdt4.arabicBN.mlf ERROR [+1232] NumParts: Cannot find word #(tdAxl in dictionary FATAL ERROR - Terminating program HLEd |
some of these words ended in \) in the mlf, which was screwing with how it appears in the dictionary. I took out the \) in the mlf, now have to make sure everything has its correct entry in the prondict. |
HLEd -d train.prondict -i train.monophone.mlf mkphones0.led tdt4.arabicBN.mlf ERROR [+6550] LoadHTKLabels: Junk at end of HTK transcription FATAL ERROR - Terminating program HLEd |
Add -T 1 to the command line. Where it stops, look in the .mlf file for that transcription. There may be a blank line or something kooky in it. This will help you find a lot of the errors that HLEd comes up with. |
HCopy -C configall -S wav2mfcc.scp ERROR [+6270] OpenParmChannel: Cannot read parameterised WAV data ERROR [+6313] OpenAsChannel: OpenParmChannel failed ERROR [+6316] OpenBuffer: OpenAsChannel failed ERROR [+1050] OpenParmFile: Config parameters invalid FATAL ERROR - Terminating program HCopy |
moved the HCopy configurations out of configall and into their own configuration file without the HCopy: prefixes |
HCopy -C confighcopy -S wav2mfcc.arabicBN.scp -T 1 data2/20000610_0330_0430_voa_arb_spl0.wav -> data2/20000610_0330_0430_voa_arb_spl0.mfcc ERROR [+6251] Input file is not in RIFF format ERROR [+6213] OpenWaveInput: Get[format]HeaderInfo failed ERROR [+6313] OpenAsChannel: OpenWaveInput failed ERROR [+6316] OpenBuffer: OpenAsChannel failed ERROR [+1050] OpenParmFile: Config parameters invalid FATAL ERROR - Terminating program HCopy |
seems to work if I put a single file on the command line couldn't figure out the problem, but it worked when I used a different computer maybe it's a 64-bit vs 32-bit problem? |
HCompV -C src/ConfigHVite -f 0.01 -v 0.01 -m -S lists/train.plp.list -M hmm0 proto/hmm0/prototype_base ERROR [+7032] FreezeOptions: vecSize not set ERROR [+5105] AllocBlock: Cannot allocate block data of 4294967288 bytes FATAL ERROR - Terminating program HCompV |
Was using the wrong hmm0/prototype; make sure it has the appropriate lines at the top (how the MFCCs were defined, E_Z_A_D etc, means of one, variances of zero |
HCompV -C src/ConfigHVite -f 0.01 -v 0.01 -m -S lists/train.plp.list -M hmm0 proto/hmm0/prototype ERROR [+7031] GetTransMat: Bad Trans Mat Sum in Row 3 HMM Def Error: GetTransMat failed at line 40/col 14/char 1028 in proto/hmm0/prototype ERROR [+7050] HMError: ERROR [+7032] LoadHMMSet: GetHMMDef failed ERROR [+2028] Initialise: LoadHMMSet failed FATAL ERROR - Terminating program HCompV |
In the prototype file, at the matrix, the copy and paste had split up the lines, so the rows did not add up to one. Make sure each row fits on a single line. |
HCompV -C configall -T 1 -A -D -m -M hmm0 -f 0.01 -S train_mfcc.list hmm0/prototype ERROR [+5050] ReadConfigFile: = expected line 1/col 8/char 7 in configall ERROR [+5020] InitShell: ReadConfigFile failed on file configall ERROR [+2000] HCompV: InitShell failed FATAL ERROR - Terminating program HCompV |
If the first column of the config file lists the program name (HVite, HCopy, etc), make sure there is a colon after the name. HCopy: TARGETKIND=MFCC_0_D_A Also make sure any '#' for comments come at the beginning of the line, not the second column. |
HCompV -c ConfigHVite -T 1 -A -D -m -M hmm0 -f 0.01 -S train_mfcc.list hmm0/prototype No HTK Configuration Parameters Set HCompV: Computing side based cepstral mean ..... ERROR [+2039] HCompV: AccGenUtt: speaker pattern matching failure on file: hmm0/prototype |
The -c needs to be -C, or else the config file isn't read. |
HCompV -C ConfigHVite -T 1 -A -D -m -M hmm0 -f 0.01 -S train_mfcc.list hmm0/prototype ERROR [+2050] CheckData: Parameterisation in ./20001001_10.mfcc is incompatible with hmm hmm0/prototype |
In hmm0/prototype, change USER to MFCC_0_D_A (when HCopy is run with MFCC_0 as the TARGETKIND |
HCompV -C ConfigHVite -T 1 -A -D -m -M hmm0 -f 0.01 -S train_mfcc.list hmm0/prototype ERROR [+2050] CheckData: Vector size in /data/data3/bromberg/fisher/segmented/fla_0069_122.mfcc[39] is incompatible with hmm hmm0/proto[13] |
In the first line of hmm0/proto, which you need to create by hand in order to run HCompV, make sure the vecSize is the same as the size of the mfccs. Here its saying that the mfcc has 39 dimensions but the proto only calls for 13. Here is a sample script for making the proto file. |
HCompV -A -T 1 -S trainsets/training-extfiles0 -l lineObservations -I labels.mlf -o lineObservations -m -M models/hmm0.0 hmmdefs/version1-hmm-top-23vec Calculating Fixed Variance HMM Prototype: hmmdefs/version1-hmm-top-23vec Segment Label: lineObservations Num Streams : 1 UpdatingMeans: Yes Target Direct: models/hmm0.0 *** stack smashing detected ***: HCompV terminated |
HTK is 32-bit program. Install GCC 3.4 for it to run it on a 64 bit machine. .. otherwise some part works / some gets stack overflow. |
HERest -C src/ConfigHVite -I lists/train.phonemlf -t 250.0 150.0 1000.0 -S train.mfcc.list -H hmm0/macros -H hmm0/hmmdefs -M hmm1 lists/monophones1 ERROR [-7324] StepBack: File ... bad data or over pruning |
Possible problems include corrupt mfcc, non-matching or non-existent labels. In this case, I had to re-calculate the mean & variance for the prototype hmm using only 1/2 the data, and the problem went away. If every file is considered bad data, you may have derived the features wrong. Go back to HCopy and check the parameters (config file). |
HERest -C src/ConfigHVite -I lists/train.phonemlf -t 250.0 150.0 1000.0 -S train.mfcc.list -H hmm0/macros -H hmm0/hmmdefs -M hmm1 lists/monophones1 Saving hmm's to dir hmm1 ERROR [+7031] PutTransMat: Row 4 of transition mat sum = 1.064684 FATAL ERROR - Terminating program HERest |
Too much data. Use the -p option, splitting the input and processing over several machines, then doing a separate HERest pass with -p 0 to accumulate the accumulators. Or, as above, use a smaller portion of the data. Also, make sure that the file durations are spread evenly across lists. Don't put all the long files together, mix them up with short ones. |
HERest -C src/ConfigHVite -I lists/train.phonemlf -t 250.0 150.0 1000.0 -S lists/train.plp.list -H hmm0/macros -H hmm0/hmmdefs -M hmm1 lists/monophones1 ERROR [+5010] InitSource: Cannot open source file hmm0/macros ERROR [+7010] LoadAllMacros: Can't open file ERROR [+5010] InitSource: Cannot open source file hmm0/hmmdefs ERROR [+7010] LoadAllMacros: Can't open file ERROR [+7050] LoadHMMSet: Macro name expected ERROR [+2321] Initialise: LoadHMMSet failed FATAL ERROR - Terminating program HERest |
Need to make 'macros' file in hmm0 directory. Copy first few lines of the prototype into macros, then append to it the vFloors file. |
HERest -C src/ConfigHVite -I lists/train.phonemlf -t 250.0 150.0 1000.0 -S lists/train.plp.list -H hmm0/macros -H hmm0/hmmdefs -M hmm1 lists/monophones1 ERROR [+5010] InitSource: Cannot open source file hmm0/hmmdefs ERROR [+7010] LoadAllMacros: Can't open file ERROR [+7050] LoadHMMSet: Macro name expected ERROR [+2321] Initialise: LoadHMMSet failed FATAL ERROR - Terminating program HERest |
Need to manually create hmmdefs file. From htkbook: "...hmmdefs containing a copy for each of the required monophone HMMs is constructed by manually copying the prototype and relabeling it for each required monophone (including sil)." Use the build_hmmdefs.py script. Add another copy of the hmm at the bottom with the label 'sil'. |
HERest -C src/ConfigHVite -I lists/all.phonemlf -t 250.0 150.0 1000.0 -S lists/train.plp.list -H hmm0/macros -H hmm0/hmmdefs -M hmm1 lists/monophones1 Pruning-On[250.0 150.0 1000.0] ERROR [+6510] LOpen: Unable to open label file /scratch/ilana/wsj/data/WSJ0/SI_TR_S/01G/01GC020X.lab FATAL ERROR - Terminating program HERest |
The label file names in the all.phonemlf file were not in all caps. Changed the script that made the word-mlf file to have the filenames in all caps, then HLEd does the phone-mlf correctly. |
HERest -A -C configall -p 3 -I train.monophone.mlf -S train.list3 -t 250.0 150.0 1000.0 -H hmm0/hmmdefs -M hmm1 monophones ERROR [+6510] LOpen: Unable to open label file /data/data3/fisher/segmented/fla_0069_122.lab FATAL ERROR - Terminating program HERest |
In the mlf file, the filenames (utterance names) did not begin with */, so they couldn't be matched to the filenames in train.list3. Make sure filenames in the mlf begin with */ and are wrapped in quotation marks. |
HERest -C src/ConfigHVite -I lists/all.phonemlf -t 250.0 150.0 1000.0 -S lists/train.plp.list -H hmm0/macros -H hmm0/hmmdefs -M hmm1 lists/monophones1 Pruning-On[250.0 150.0 1000.0] WARNING [-7325] LoadUtterance: No labels in file /scratch/ilana/wsj/data/WSJ0/SI_TR_S/01G/01GO031F.lab in HERest Segmentation fault |
Rework prompts2mlf_word.py to not let a file begin with '.'; redo word and phone mlfs. |
HERest -C src/ConfigHVite -I lists/all.phonemlf -t 250.0 150.0 1000.0 -S lists/train.plp.list -H hmm0/macros -H hmm0/hmmdefs -M hmm1 lists/monophones1 Pruning-On[250.0 150.0 1000.0] ERROR [+7011] SaveHMMSet: Cannot create MMF file hmm1/macros |
mkdir hmm1 |
HERest -C ConfigHVite -I 20001001_1.monophone.mlf -t 250.0 150.0 1000.0 -S train_plp.list -H hmm0/macros -H hmm0/hmmdefs -M hmm1 monophones HMM Def Error: GetToken: Symbol expected at line 1/col 4/char 3 in hmm0/macros ERROR [+7050] HMError: HMM Def Error: GetOptions: GetToken failed at line 1/col 5/char 4 in hmm0/macros ERROR [+7050] HMError: HMM Def Error: LoadAllMacros: GetOptions Failed at line 1/col 0/char -1 in hmm0/macros ERROR [+7050] HMError: HMM Def Error: LoadAllMacros: Macro sym expected at line 1/col 0/char -1 in hmm0/hmmdefs ERROR [+7050] HMError: ERROR [+7050] LoadHMMSet: Macro name expected ERROR [+2321] Initialise: LoadHMMSet failed FATAL ERROR - Terminating program HERest |
hmmdefs file is screwy, the line ~h 'aa' needs to come before the BEGINHMM line. |
HERest -C ConfigHVite -I 20001001_1.monophone.mlf -t 250.0 150.0 1000.0 -S train_plp.list -H hmm0/macros -H hmm0/hmmdefs -M hmm1 monophones Pruning-On[250.0 150.0 1000.0] ERROR [+6510] LOpen: Unable to open label file 20001001_1.plp.lab FATAL ERROR - Terminating program HERest |
The filenames have to be matching within the mlf files and the individual names of the pfiles. xxx.lab and xxx.pf, no variations. |
HERest -A -C configall -p 1 -I train.monophone.mlf -S train.list1 -t 250.0 150.0 1000.0 -H hmm0/hmmdefs -M hmm1 monophones ERROR [+5105] AllocBlock: Cannot allocate block data of 5000000 bytes FATAL ERROR - Terminating program HERest |
One of the training files is too large for the system to process. Rerun the HERest command with -T 1, see what file it fails on, remove it from train.list1, and try again. (Alternatively split up that file and its transcript and replace the original file with its splits in the training list and the mlf.) |
HHEd -H hmm4/macros -H hmm4/hmmdefs -M hmm5 sil.hed lists/monophones1 ERROR [+7030] GetHMMDef: Trans Mat Dimensions not 3 x 3 HMM Def Error: LoadAllMacros: GetHMMDef failed at char 188656 in hmm4/hmmdefs ERROR [+7050] HMError: ERROR [+7050] LoadHMMSet: Macro name expected ERROR [+2628] Initialise: LoadHMMSet failed FATAL ERROR - Terminating program HHEd |
Add sp to monophones 1; To make hmmdefs4, use this script: sil2sp.pl, which I got from this htk tutorial website. |
HHEd -H hmm4/macros -H hmm4/hmmdefs -M hmm5 src/sil.hed lists/monophones1 WARNING [-2631] EditTransMat: No trans mats to edit! in HHEd |
Was using wrong monophones list; must use the one updated with 'sp' |
HERest -A -C configall -I train.monophone.sp.mlf -S train.list -t 250.0 150.0 1000.0 -H hmm5/hmmdefs -M hmm5 monophones.sp Pruning-On[250.0 150.0 1000.0] ERROR [+7332] Create Insts: Cannot have Tee models at start or end of transcription FATAL ERROR - Terminating Program HERest |
There is an 'sp' short pause as the last symbol before '.' in the mlf. In a previous step there was an HLEd command with a set of commands in a file like 'mkphones1.led'. Make sure 'IS sil sil' is in that .led file, which puts the 'sil' at beginning and end of each utterance. |
HERest -C src/ConfigHVite -I lists/all.sp.phonemlf -t 250.0 150.0 1000.0 -S lists/train.plp.list -H hmm5/macros -H hmm5/hmmdefs -M hmm6 lists/monophones1.sp Pruning-On[250.0 150.0 1000.0] ERROR [+7332] CreateInsts: Cannot have Tee models at start or end of transcription FATAL ERROR - Terminating program HERest |
Recreate phone mlf to have 'sil' before and after each utterance; use "IS sil sil" in the .led file for HLed; if it still doesn't work, find by hand the utterances that end in 'sp' and add 'sil' before the period; or use the python command line to fix it. |
HVite -l '*' -o SWT -b SILENCE -C ConfigHVite -a -H hmm7/macros -H hmm7/hmmdefs -i 20001001_1.realigned.monophone.mlf -m -t 250.0 -y lab -I 20001001_1.mlf -S train_plp.list prondict.sort.sp monophones ERROR [+6510] LOpen: Unable to open label file 20001001_1.lab FATAL ERROR - Terminating program HVite |
name of .lab file in 20001001_1.mlf (word mlf) was wrong |
HVite -l '*' -o SWT -b SILENCE -C ConfigHVite -a -H hmm7/macros -H hmm7/hmmdefs -i 20001001_1.realigned.monophone.mlf -m -t 250.0 -y lab -I 20001001_1.mlf -S train_plp.list prondict.sort.sp monophones nothing appears in new transcription |
Trying to realign the transcription using the current hmms. This may be the fault of me not training with enough data. Try just copying the original monophone transcript to realigned.monophone.mlf and continue |
HERest -C configall -I 20001001_1-10.realigned.monophone.mlf -t 250.0 150.0 1000.0 -S train_mfcc.list -H hmm7/macros -H hmm7/hmmdefs -M hmm8 monophones Pruning-On[250.0 150.0 1000.0] ERROR [+6510] LOpen: Unable to open label file 20001001_5.lab FATAL ERROR - Terminating program HERest |
The HVite process did not create a label file for every utterance; some had no tokens surviving, including file 5. Need to go back to HVite and change some parameters to make sure it can get through all utterances. For instance, change the beam searching parameters with the -t flag (htkbook pg 301) |
HERest -C configall -I train.realigned.monophone.mlf -t 250.0 150.0 1000.0 -p 0 -H hmm7/macros -H hmm7/hmmdefs -M hmm8 hmm8/HER1.acc hmm8/HER2.acc hmm8/HER3.acc hmm8/HER4.acc hmm8/HER5.acc hmm8/HER6.acc ERROR [+7060] InitHMMSet: Expected newline after 2'th HMM ERROR [+2321] Initialise: MakeHMMSet failed FATAL ERROR - Terminating program HERest |
forgot to put 'monophones' on the command line before the list of .acc files |
HVite -T 1 -C configall -H hmm9/macros -H hmm9/hmmdefs -S train_mfcc.list -l '*' -i recog_mono2/monophones.mlf -o S -w wdnet -p 0.0 -s 5.0 prondict.sort.final monophones Read 37 physical / 37 logical HMMs WARNING [-8520] CreateSEIndex: No transitions to state 5 in HVite WARNING [-8520] CreateSEIndex: No transitions to state 5 in HVite Read lattice with 859 nodes / 1713 arcs Created network with 6391 nodes / 7245 links |
I'm doing recognition at the hmm9 stage as part of debugging. There are a few hmms in hmm8/hmmdefs and hmm9/hmmdefs that have no transition from state 4 to state 5, including 'O' and 'silst'. This is a problem. It start occurring after realignment. So for the two iterations of HERest after realignment, use -u mv |
HHEd -B -H hmm9/macros -H hmm9/hmmdefs -M hmm10 src/mktri.hed lists/monophones1.sp ERROR [+2635] FindBaseModel: Cannot Find HMM sl in Current List FATAL ERROR - Terminating program HHEd |
Found and removed 'sl' from lists/triphones1 |
HHEd -T 1 -H hmm9/hmmdefs -M hmm10 mktri.hed monophones HHEd 34/34 Models Loaded [5 states max, 1 mixes max] CL triphones Cloning current hmms to produce new set {(*-. Error ) expected ERROR [+7230] EdError: item list parse error FATAL ERROR - Terminating program HHEd |
Because '.' is in the monophones list, the first triphone code in mktri.hed is invalid. remove it. |
HERest -B -A -C configall -s stats -p 0 -I train.triphone.cw.mlf -t 250.0 150.0 1000.0 -H hmm11/macros -H hmm11/hmmdefs -M hmm12 triphones hmm12/HER1.acc hmm12/HER2.acc Pruning-On[250.0 150.0 1000.0] ERROR [+7191] Infinite WtAcc! (or) ERROR [+7191] Infinite MuAcc! FATAL ERROR - Terminating program HERest |
Comes up in the accumulation process when doing a split re-estimation. WtAcc due to a row sum error on the transition matrix.(?) Tried: -u mv in the command, gave me MuAcc instead Tried: using files less than a minute in length Tried: splitting data up into more parallel sections (4) Tried: remove -B from the combining step in HERest, making the resulting hmm text form rather than binary. This worked, but I don't know why... |
HHEd -H hmm12/macros -H hmm12/hmmdefs -M hmm13 tree.hed triphones ERROR [+2662] AssignStructure: cannot find tree for U-r+sil state 5 FATAL ERROR - Terminating program HHEd |
I'm using 5 middle states but tree.hed only has TB lines for 3 middle states. Add more TB lines to tree.hed for states 5 & 6 |
HHEd -B -H hmm12/hmmdefs -M hmm13 tree.hed triphones ERROR [+2662] AssignStructure: cannot find tree for t2-ay+D2 state 2 FATAL ERROR - Terminating program HHEd |
One of the phonemes in the triphone listed has no indication of how to cluster it in tree.hed. Remove it from the prondict and start over (with a shortened monophone list), or remove it from fulllist and the HHEd command will run. If you have one like this you probably have a few, look carefully. |
HHEd -B -H hmm12/macros -H hmm12/hmmdefs -M hmm13 src/tree.hed lists/triphones1 > log ERROR [+2662] AssignStructure: cannot find tree for ax-sp+d state 2 FATAL ERROR - Terminating program HHEd |
Recreate tree.hed using local monophone list, mkclscript from tutorial, then add to that QS part of tree.hed |
HHEd -H hmm12/macros -H hmm12/hmmdefs -M hmm13 tree.hed triphones ERROR [+2662] FindProtoModel: no proto for z-sp+A in hSet FATAL ERROR - Terminating program HHEd |
In the cross-word triphone models, the sp causes problems, so remove it from the monophone list, from the extra triphones, from tree.hed |
HERest -B -C ConfigHVite -I 20001001_1.triphone.mlf -t 250.0 150.0 1000.0 -S train_plp.list -H hmm13/macros -H hmm13/hmmdefs -M hmm14 triphones ERROR [+5010] InitSource: Cannot open source file Q-n+A ERROR [+7010] LoadHMMSet: Can't find file ERROR [+2321] Initialise: LoadHMMSet failed FATAL ERROR - Terminating program HERest |
The state-tying actually caused some states to be tied, meaning they get renamed. This is shown in the tiedlist created in the previous step with HHEd and CO "tiedlist" at the end of tree.hed. In the tiedlist output file, there are two columns in some places; the second column names the new label for the hmm in the first. It means those two are tied. So Q-n+A is tied to another triphone and thereby renamed. REPLACE TRIPHONES WITH TIEDLIST ON THE COMMAND LINE. |
HERest -B -C configall -I train.triphone.mlf -t 250.0 150.0 1000.0 -S train.realigned.list -H hmm13/macros -H hmm13/hmmdefs -M hmm14 tiedlist ERROR [+7231] InitSource: Cannot open source file y-l-A FATAL ERROR - Terminating program HERest |
There is a triphone that HERest is trying to reestimate that does not appear in the tiedlist. Recreate the fulllist (all possible triphones) and redo the HHEd step for decision tree tying etc. |
HERest -T 1 -D -A -C configall -p 1 -I train.triphone.mlf -S train.mfcc.norm.list -t 250.0 150.0 1000.0 -H hmm15/hmmdefs -M hmm16 tiedlist HERest ML Updating: Transitions Means Variances Parallel-Mode[1] System is SHARED 51987 Logical/15201 Physical Models Loaded, VecSize=39 1 MMF input files Pruning-On[250.0 150.0 1000.0] Processing Data: fla_0130_96.mfcc; Label fla_0130_96.lab Utterance prob per frame = -5.125980e+01 Processing Data: fla_0530_22.mfcc; Label fla_0530_22.lab ERROR [+7321] CreateInsts: Unknown label m+H FATAL ERROR - Terminating program HERest |
add a line to the mktri.led file that has 'NB sp' or 'NB garbage' or 'NB whatever' for whatever monophone for which you don't want the biphone context to be made. Go back and remake the triphone transcript, then try the re-estimation again. |
HHEd -A -H mix_moreA/hmmdefs -M mix_moreA 10.hedscript tiedlist WARNING [-2637] HeaviestMix: mix 4 in n2-O+sh2 has v.small gConst [-200000045056.000000] in HHEd WARNING [-2637] HeaviestMix: mix 1 in n2-O+sh2 has v.small gConst [-170000023552.000000] in HHEd WARNING [-2637] HeaviestMix: mix 3 in n2-O+sh2 has v.small gConst [-109999996928.000000] in HHEd WARNING [-2637] HeaviestMix: mix 4 in n2-O+sh2 has v.small gConst [-140000002048.000000] in HHEd ERROR [+2697] HeaviestMix: heaviest mix is defunct! FATAL ERROR - Terminating program HHEd |
Trying to increase the number of Gaussian mixtures for each hmm at the end of training, incrementing by 2 each time. From htkbook: "Defunct mixture components can be prevented by setting the -w option in HERest so that all mixture weights are floored to some level above MINMIX." |
HVite -H hmm15/macros -H hmm15/hmmdefs -S lists/dt.list -l '*' -i recog/dt.out.mlf -w wdnet -p 0.0 -s 5.0 lists/allwords.prons.dict.final lists/tiedlist ERROR [+8251] ReadLattice: Word worrisome not in dict ERROR [+3210] DoAlignment: ReadLattice failed FATAL ERROR - Terminating program HVite |
Made sure vocab, wordnet were all uppercase; dictionary is all uppercase; |
HVite -H hmm15/macros -H hmm15/hmmdefs -S lists/dt.list -l '*' -i recog/dt.out.mlf -w wdnet.upper -p 0.0 -s 5.0 lists/allwords.prons.dict.final lists/tiedlist ERROR [+8251] ReadLattice: Word -PAU- not in dict ERROR [+3210] DoAlignment: ReadLattice failed FATAL ERROR - Terminating program HVite | add it to the pronunciation dictionary |
HVite -H hmm15/macros -H hmm15/hmmdefs -S lists/dt.list -l '*' -i recog/dt.out.mlf -w wdnet.upper -p 0.0 -s 5.0 lists/allwords.prons.dict.addrecog lists/tiedlist WARNING [-8221] InitPronHolders: Total of 77 duplicate pronunciations removed in HVite ERROR [+8231] GetHCIModel: Cannot find hmm [???-]IY[+???] FATAL ERROR - Terminating program HVite |
Change dictionary and wdnet to all lowercase |
HVite -T 1 -C src/ConfigHVite -H hmm15/hmmdefs -H hmm15/macros -S lists/dt.list -i recog/dt.out.mlf -o S -w wdnet.lower -p -10.0 -s 15.0 -t 450.0 250.0 40000.0 lists/allwords.rons.dict.addrecog.lower lists/tiedlist > recog.log ERROR [+8250] ReadLattice: Premature end of lattice file before header ERROR [+3210] DoAlignment: ReadLattice failed FATAL ERROR - Terminating program HVite |
Go back into wdnet.lower and uppercase the first line and the J,I,W,S,L etc |
HVite -T 1 -C src/ConfigHVite -H hmm15/hmmdefs -H hmm15/macros -S lists/dt.list -i recog/dt.out.mlf -o S -w wdnet.lower -p -10.0 -s 15.0 -t 450.0 250.0 40000.0 lists/allwords.rons.dict.addrecog.lower lists/tiedlist > recog.log ERROR [+8231] GetHCIModel: Cannot find hmm [l-]e[+sh] FATAL ERROR - Terminating program HVite |
Changed pronunciation of [inhalation] to l ey sh There is a monophone somewhere in the dictionary, or in the monophone set, that is not represented as an hmm (try "cat hmm0/hmmdefs | grep '~h'" to see what is represented). You may need to either change pronunciations in the dictionary to eliminate barely-used monophones or retrain with all of the monophones intact. It's possible if you generated the monophone list from the monophone transcript that some monophones in the prondict were left out, b/c they never occurred in the first pronunciation of any word. Try regenerating the monophone list from the dictionary using shell scripting instead of HLEd. |
HVite -T 1 -C src/ConfigHVite -H hmm15/hmmdefs -H hmm15/macros -S lists/dt.list -i recog/dt.out.mlf -o S -w wdnet.lower -p -10.0 -s 15.0 -t 450.0 250.0 40000.0 lists/allwords.rons.dict.addrecog.lower lists/tiedlist > recog.log ERROR [+6313] OpenParmChannel: cannot read HTK Header in File /u/drspeech/data/WSJ0/SI_DT_05/050/050A0503.nst ERROR [+6313] OpenAsChannel: OpenParmChannel failed ERROR [+6316] OpenBuffer: OpenAsChannel failed ERROR [+3250] ProcessFile: Config parameters invalid FATAL ERROR - Terminating program HVite | Changed dt.list to dt.plp.list |
HVite -z lat -l $expname -C ../configall -t 150.0 -A -D -T 1 -w $expname.htk.lm -s 12.0 -p -10.0 -H ../hmmdefs.16 -S ../dev.mfcc0.list1 prondict.norm8.sort.sp ../tiedlist ERROR [+8231] GetHCIModel: Cannot find hmm [u-]n[+???] FATAL ERROR - Terminating program HVite |
Haven't figured this one out. Can't find a pronunciation with fishy phonemes as mentioned. HDecode has no problem with all of the same inputs except for ARPA-based lm, and I don't see anything wrong with the htk-lattice-lm. So, I dunno. |
HHEd -B -H hmm15/macros -H hmm15/hmmdefs -M hmm16 src/train_mix_inc_2.hed lists/train+cv.triphonemlf ERROR [+7036] CreateHMM: multiple use of logical HMM name sp ERROR [+7060] InitHMMSet: Error in CreateHMM ERROR [+2628] Initialise: MakeHMMSet failed FATAL ERROR - Terminating program HHEd |
|
Reading dictionary from diss/lib/myprondict ERROR [+8050] ReadDict: Probability malformed 2 ERROR [+8013] ReadDict: Dict format error ERROR [+9999] Initialise: ReadDict failed FATAL ERROR - Terminating program HDecode.long |
problems in the pronunciation dictionary: quotations and double quotes need backslash brackets possibly need backslash narrow down problem by reducing prondict to only a few lines and gradually adding until the error comes up one of the last pronunciations has a non-existent phoneme (2), change it. |
HDecode.long -z lat -l decodeLCA_nonums_wordLM -C configall -t 150.0 -A -D -T 1 -w lev.alltext.word.lm -s 12.0 -p -10.0 -H mix_moreA/hmmdefs -S lev.dev.mfcc.list4 levtrain.prondict.ver3 tiedlist Reading dictionary from levtrain.prondict.ver3 Reading acoustic models... Read 4163 physical / 230643 logical HMMs ERROR [+9999] HLVNet: no model label for phone (uw-gar+gar) FATAL ERROR - Terminating program HDecode.long |
I have a 'gar'bage model that is like sp, should not belong to any triphones. In HDecode, only the phonemes associated with start and/or endnode are allowed to be monophone-only. Go back and add gar triphones to full_list, remake the tiedlist. Might need to add some info for gar to tree.hed. Re-estimate from there forward. |
HDecode.long -z lat -l * -C ConfigHVite -t 150 -A -D -T 1 -w 20001001_1.lm -s 12.0 -p -10.0 -H hmm15/hmmdefs -S train_plp.list prondict.sort tiedlist ERROR [+4019] HDecode: beam width expected FATAL ERROR - Terminating program /u/drspeech/opt/htk-3.4/i586-linux/bin/HDecode.long |
The value after the -t flag must be a float. Change to 150.0 |
HDecode.long -z lat -l * -C ConfigHVite -t 150.0 -A -D -T 4 -w 20001001_1.lm -s 12.0 -p -10.0 -H hmm15/hmmdefs -S train_plp.list prondict.sort tiedlist ERROR [+9999] HDecode: cannot find STARTWORD ' FATAL ERROR - Terminating program /u/drspeech/opt/htk-3.4/i586-linux/bin/HDecode.long |
add |
HDecode.long -z lat -l * -C ConfigHVite -t 150.0 -A -D -T 4 -w 20001001_1.lm -s 12.0 -p -10.0 -H hmm15/hmmdefs -S train_plp.list prondict.sort tiedlist ERROR [+9999] HDecode: cannot find file 'sp' FATAL ERROR - Terminating program /u/drspeech/opt/htk-3.4/i586-linux/bin/HDecode.long |
add sp sil to the end of the prondict |
HDecode.long -z lat -l decodeA -C ../configall -t 150.0 -A -D -T 1 -w mix2.unk.knd.lm -s 12.0 -p -39.0 -H ../hmmdefs.16 -S ../dev.mfcc0.list4 prondict.expand ../tiedlist FATAL ERROR - Terminating program HDecode.long ERROR [+5010] InitSource: Cannot open source file f-uw+x ERROR [+7010] LoadHMMSet: Can't find file ERROR [+4128] Initialise: LoadHMMSet failed |
There is a mismatch between the hmms that are defined in the hmmdefs file and those that are listed in the tiedlist. One or the other needs to change, probably the tiedlist. This may involve going back far enough to re-create the hmms used in the last HHEd command, so to recreate the tiedlist. |
HDecode.long -z lat -l * -C ConfigHVite -t 150.0 -A -D -T 4 -w 20001001_1.lm -s 12.0 -p -10.0 -H hmm15/hmmdefs -S train_plp.list prondict.sort tiedlist WARNING [-9999] no token survived to sent end! in HDecode.long Segmentation fault |
This is the model I built on a single sound file, so maybe that's the right answer... |
ERROR [+9999] HLVNet: no model label for phone (.-q+r) FATAL ERROR - Terminating program /u/drspeech/opt/htk-3.4/i586-linux/bin/HDecode.long |
Remove '. .' from the pronunciation dictionary |
HDecode.long -z lat -l decodeA -C configall -t 150.0 -A -D -T 1 -w p.3grams.lm -s 12.0 -p -10.0 -H mix_moreA/hmmdefs.16 -S dev.mfcc0.list3 prondict.norm8.sort tiedlist ERROR [+9999] HLVNet: no model label for phone (x-sil+S) FATAL ERROR - Terminating program HDecode.long |
It shouldn't be looking for a triphone with 'sil' in the middle. Search for 'sil' in the pronunciation dictionary; the only words it should serve as pronunciation for are |
ERROR [+9999] HDecode: Incompatible parm kinds MFCC_0 vs. MFCC_D_A_0 FATAL ERROR - Terminating program /u/drspeech/opt/htk-3.4/i586-linux/bin/HDecode.long |
Changed the format of the hmms to mfcc_d_a_0 even though the original files were made into MFCC_0. They've gotta be the same. Use MFCC_D_A_0 in the hcopy config file. |
Reading dictionary from diss/lib/myprondict Reading acoustic models...Read 26745 physical / 250049 logical HMMs ERROR [+9999] HLVNet: no model label for phone (sil-}+w) FATAL ERROR - Terminating program HDecode.long |
There are still some labels in the pronuncation dictionary that do not have defined acoustic models (}). Change those labels, which may have come in through the pronunciation-building script. |
ERROR [+8113] ReadARPAngram: failed reading lm prob at char 1283900 in diss/data/language_model/fsms.4grams.64Kvocab.lm | This error can be reproduced by having the wrong number of ngrams present in the lm file as compared to the number defined at the top of the file. Make sure the LM and prondict have the same encoding. Make sure all quotes and double quotes have a backslash Make sure the lm and prondict contain \ |
WARNING [-8100] ReadARPAngram: unseen word 'إسأ' in ngram in HDecode.long | Words in lm not present in pronunciation dict. Write a script to find them and add them in, being sure to resort the pronunciation dictionary afterwards. |
ReadNGrams: 1827th 2Grams out of order | The bigrams are not in alphabetical order. |
HLRescore -n $lm -f -y crec -r 10.0 -t 150.0 -s 20.0 -p -42.0 -C configall -A -D -T 1 -S unconstrained.list $prondict.expand Reading LM from mix1.unk.knd.lm ERROR [+8150] ReadNGrams: 577308th 2Grams out of order FATAL ERROR - Terminating program HLRescore |
Nothing seems out of place in the LM which was made and not messed with. LM worked for HDecode. |
HLRescore -n $lm -f -y crec -t 150.0 -s 12.0 -p -10.0 -C configall -A -D -T 1 $prondict $lattice WARNING [-9999] word 0 not in LM wordlist in HLRescore HLRescore: HLat.c:415: LatTopSort: Assertion `time+1 == lat->nn' failed. |
Got rid of the second error (LatTopSort) by determinizing, minimizing, and topologically sorting the fsm before converting to pfsg & htklat. Make sure that NULL (or whatever you've substituted for NULL in the lattice) exists in both the pronunciation dictionary and the language model. |
HLRescore -n $lm -f -y crec -t 150.0 -s 12.0 -p -10.0 -C configall -A -D -T 1 $prondict $lattice ERROR [+8250] ReadLattice: Premature end of lattice file before header ERROR [+4013] HLRescore: can't read lattice FATAL ERROR - Terminating program HLRescore |
In the lattice, change 'NODES' to 'N' and 'LINKS' to 'L'. |
HLRescore -n $lm -f -y crec -t 150.0 -s 12.0 -p -10.0 -C configall -A -D -T 1 $prondict $lattice ERROR [+8251] ReadLattice: Word bEd:bEd not in dict ERROR [+4013] HLRescore: can't read lattice FATAL ERROR - Terminating program HLRescore |
Needed to use transducer=1 to get the pfsg to print correctly with fsm-to-pfsg, but now the format is messing up HLRescore. Either change to transducer=0 in fsm-to-pfsg, or, use sed to change each term:term to term. |
HResults -I reference.mlf /dev/null decoded.mlf ERROR [+6550] LoadHTKList: Label Name Expected FATAL ERROR - Terminating program HResults |
In the reference.mlf file, there exists either a blank line, or a digit without backslash or quotes, or something else unpalatable to HResults. Use HResults -f to figure out which utterance it's in (the one _after_ the last one listed), and fix it. For instance, put quotations around a number or backslash a quote, etc. |
HResults -I reference.mlf /dev/null hypothesis.mlf ERROR [+6570] Get LabelList: n[1] > numLists[0] FATAL ERROR - Terminating program HResults |
Run the command again with -f to show full results. Look in the reference mlf file at the utterance _after_ the last one listed before the error shows up. It's empty. Put something there or remove it (must be removed from reference). |
HResults -I reference.mlf /dev/null hypothesis.mlf ERROR [+6510] LOpen: Unable to open label file NBCTV_MORNING_20070111.lab FATAL ERROR - Terminating program HResults |
One of the utterances in the hypothesis mlf does not have a corresponding utterance in the reference mlf. Either it's missing entirely or the names don't match, check spelling and capitalization of the filenames in the two mlfs. |
HERest -A -D -T 10 -C configall -C hmmadapt6-1/config_adapt -S adapt6.list -I train.triphone.new.mlf -H hmmadapt6-1/hmmdefs.16 -H hmmadapt6-1/glob -K hmmadapt6-2 mllr -u a tiedlist ERROR [+999] Components missing from Base Class list (4630 74080) ERROR [+999] BaseClass check failed FATAL ERROR - Terminating program HERest |
Trying to do adaptation. Built regression tree using instructions here. But because my hmm definitions were incremented to 16 mixes, I had to change the last line of the global file to: CLASS 1 {*.state[2-4].mix[1-16]} |
HERest -A -D -T 10 -C configall -C hmmadapt6-1/config_adapt -S adapt6.list -I train.triphone.new.mlf -H hmmadapt6-1/hmmdefs.16 -H hmmadapt6-1/glob -K hmmadapt6-2 mllr -u a tiedlist ERROR [+999] Output xform mask *.%%% does not match filename data2/20001220_1530_1600_NTV_ARB/20001220_1530_1600_NTV_ARB_3.mfcc FATAL ERROR - Terminating program HERest |
Added -h data2/*/*_ARB_??.mfcc to the command line, which matches the filenames being given to HERest. Also take away 'mllr' |
HERest -C configall -C hmmadapt6-1/config_adapt -S adapt6.list -I train.triphone.new.mlf -z tmf -h data2/*/*_ARB_??.mfcc -H hmmadapt6-1/hmmdefs.16 -H hmmadapt6-1/glob -u a -K hmmadapt6-2 -M hmmadapt6-2 -d hmmadapt6-1 tiedlist ERROR [+7060] InitHMMSet: Expected newline after 1'th HMM ERROR [+2321] Initialise: MakeHMMSet failed FATAL ERROR - Terminating program HERest |
The "-h data2/*/*_ARB_??.mfcc" command is causing a bunch of filenames to come up as part of the command, and it thinks that one of those is the list of hmms. Instead, use data2/*/*, and PUT SINGLE QUOTES AROUND IT: -h 'data2/*/*' |
HERest -C configall -C config.global -S adapt6.list -I train.triphone.new.mlf -u a -z tmf -K xforms mllr1 -J classes -h 'data2/*/*' -H hmmorig/hmmdefs.16 tiedlist no output |
No errors, but no new files in xforms or anywhere else. Change the -h part of the command. Wherever there is a '%', that part will be the name of the output file (plus mllr or whatever comes after the output directory after the -K flag). So if you have two input sources or speakers, put the %% to match the characters that group each speaker's utterances, give only those utterances as input, and you'll get a transform for that speaker with an appropriate name. |
posted on
2013-07-10 17:23 ldjhust 阅读(
...) 评论(
...) 编辑 收藏