we got error: last state for word SAZ is not sp. Sorry - not implemented.

The problem is it is that I removed sp from SAZ so that it is not sil sp but sil.

If cython is not available, this package fails setup.py install with this error:

    Complete output from command python setup.py egg_info:
    Traceback (most recent call last):
      File "<string>", line 20, in <module>
      File "/mnt/compmusic/itri/jenkins/jobs/Dunya/workspace/env/src/alignment-duration/setup.py", line 62, in <module>
        cmdclass = {'build_ext': build_ext},
    NameError: name 'build_ext' is not defined

This is preventing dunya from building and running tests, and so we have currently removed it as a dependency

as well LyricsWithModels.LyricsWithModels.printWordsAndStates and LyricsWithModels.LyricsWithModels.printWordsAndStatesAndDurations can take code from LyricsParsing.expandlyrics2WordList.

onsets.OnsetDetector.OnsetDetector.parseNoteOnsetsGrTruth
and
replace also parse.TextGrid_Parsing._findBeginEndIndices

in file LyricsParsing

in function constructTransMAtrix from Decoder : # SPECIAL CASE: two last states
the else is not needed? because defineForwardTransProbs() covers it.

Related: THink how to reduce code do that # MAIN CASE and SPECIAL CASE are merged and defineForwardTransProbs() is called once only

TODO: could be done weasier with this code, and check last method in Word

grTruthWordList = testT(decoder.lyricsWithModels)

for getRefDurations : expandlyrics2WordList
for decoded done with other code.
unify these two.

solve together with HARD CODED bug

Move this file to dir for_english
/Users/joro/workspace/AlignmentDuration/src/for_makam/state_str2int_METU

means, covars, weights, should not be in the constructor of _ContinuousHMM, because they are observation probabilities.

Store the char, not only the pinynin in the mandarinSyllable class. See initi of the mandarinSyllable class here. Split chinese characters in this method and give as argument.

e.g. printWordsAndStatesAndDurations and phoneteme2StatesNetwork.
Put this repeating logic as a helper method in LyricParsing

Disable the link to the lyric player for all recordings that are not in the list of recording MBIDs with singing voice present.

happens when t < MAX_DUR.
probable reason:
phi Star stays more states than currMaxDur.

in lines starting at
if ParametersAlgo.WITH_ORACLE_PHONEMES: # oracle phonemes

https://github.com/georgid/AlignmentDuration/blob/noteOnsets/align/LyricsAligner.py#L253

in line 117 of AlignmentStep.Aligner() there is different num arguments. problem triggered by outputHTKPhoneAlignedURI = Aligner.alignOnechunk(MODEL_URI, URIrecordingWav, lyrics, URIrecordingAnno, '/tmp/', withSynthesis)
in AlignmentDuration.alignOneChunk()

this is dangerous

try to Remove

in align.LyricsWithModels.LyricsWithModels._linkToModels

This way not needed to handle specially the phoneme sp. It is now inserted as Phoneme in phonemes network in LyricsWithModels but not seen in list of words in Lyrics

hmm.Path.Path.init can haas input only hmm and read psi and phi from it

make SymbTrParser and maybe MakamScore inherit form AlignmentStep.*

make sure extracting MFCC with essentia same as damp model:

• add preempahsis (or recreate model without preemphasis )
• add cepstral mean normalization

dont use scikit learn at all, keep LyricsWIthModelsGMM class for chinese.

• test on Jingju

instead of adapting lyricsWithModels class to guyzs implementation,
adapt the viterbi search to lyricsWithModels

MEthods to change: parse lyricsWithModels and construct params for guyZ
=Decoder.Decoder._constructHMMNetworkParameters

=Decoder.Decoder.path2ResultWordList - uses stateNetwork indices

-LyricsAligner. withAnnotations and withLinks have two similar loops, think how to put it in one loop.

• test and update pycompmusic.lyricsAlign

sectionLInk object has no section object assigneed.
See in method
makam.MakamRecording.MakamRecording._loadsectionTimeStampsLinks
have a look for an example at:
makam.MakamRecording.MakamRecording._loadsectionTimeStampsAnno()

This fails in align.LyricsAligner.LyricsAligner.alignRecording:
if not hasattr(currSectionLink, 'section') or currSectionLink.section == None:

there is kimseye-specific code in
align.LyricsAligner.LyricsAligner.alignLyricsSection

remove it.

This is the SIL_TOKEN item

https://github.com/georgid/AlignmentDuration/blob/noteOnsets/src/align/LyricsAligner.py#L142

test from for_makam/lyrics_align WITH_DURATIONS = True
first

• get rid of loading htk models, and get rid of class LyricsWithModelsHTK (see issue #40)
• select recordings with no second verse
• let andres run on dunya
• optionally divide into good and bad quality

_constructHMMNetworkParameters is on 3 different places. Leave onlpy in ContinousHMM or Decoder.

here)

AlignmentEvaluation has two branches:

1. https://github.com/georgid/AlignmentEvaluation/tree/for-audio-segmentsused for makam data which is annotated per audio segment (section of 10-20 seconds)
   http://compmusic.upf.edu/turkish-makam-acapella-sections-dataset
2. https://github.com/georgid/AlignmentEvaluation/tree/master for jingju data

EXAMPLE : /Users/joro/Documents/Phd/UPF//ISTANBUL//barbaros/02_Gel_9_nakarat2.scoreDeviation

this is a problem for evaluation Accuracy::
For now WORKAROUD in AccuracyEvaluator.calcCorrect
currEndDetected = finallTsAnno
logging.warn("currEndDetected > finallTsAnno")

REASON: Munir nurretim makes shorter some notes? ??

Would be nice that the field sectionLyrics in MUsicXML Parser is type ScoreSection.
further MakamScore should have a field score in the same way symbTrParser has a field score of type MakamScore.

Check if Path makes sence. put just backtracking logic in Path

The LyricsWithModels is not needed for NeuralNEtwork, so Baseclass is used for DNN, add padded silicce method. As result:

• LyricsWithModelsBase is used for DNN. LyricsWithModelsCNN._linkTomodels and
• LyricsWithModelsBase._linkTomodels do not make sense.

maybe remove lyrics With models in general.

then reduce if statement in SecionLink.loadSmallAudioFragment()

_phonemes2stateNetwork(self):

check steps from TODO

Decoder.Decoder.duration2numFrameDuration

The code for expansion of silence word could be reduced here

No round() at
Decoder.Decoder.duration2numFrameDuration
but instead put this as mean value and round() duration times

make sure extracting MFCC with essentia same as damp model:

add preempahsis (or recreate model without preemphasis )
add cepstral mean normalization

in function detectedAlignedfileName = mlf2TabFormat(detectedWordList, URIrecordingNoExt, tokenLevelAlignedSuffix)

so eliminate RecordingSegmenter step,
do alignment for each section in a loop

• recomupte timestamps fron whole recording mfc to given sectuin( input: section.json)
• Align
• recomupte tinmestamps for whole recordings

Concatenate TextGrid annotation files for the segmented files
into one-per-recording TextGrid annotation automatically:

1. install TextGridTools version 1.4.1 (either through GitHub or pip install --upgrade tgt). If you use pip, please make sure that it is really version 1.4.1 that is installed — if you get an older version, try again.
2. QUESTION:
   Because each big audio file to which the concatenated text grid corresponds starts with n seconds of silence, then I need to insert in the beginning n seconds of silence and then concatenate TextGrids where the first starts at timestamp = n.

To do this I tried this code:

shiftTime = 51.354230

tiers_ = []
os.chdir(pathInput)
tgtURI = '/Users/joro/Documents/Phd/UPF/ISTANBUL/goekhan/02_Kimseye_2_zemin.TextGrid'

from tgt.util import shift_boundaries
tg = tgt.read_textgrid(tgtURI)

tier = tg.get_tier_by_name('words')
tierShifted = shift_boundaries(tier, shiftTime,0)

tg.add_tiers(tierShifted)

tgOutURI = pathOut + 'Kimseye.TextGrig'
tgt.write_to_file(tg, tgOutURI)

However I get this error:

in ()
21
22 tgOutURI = pathOut + 'Kimseye.TextGrig'
---> 23 tgt.write_to_file(tg, tgOutURI)

/usr/local/lib/python2.7/site-packages/tgt/io.pyc in write_to_file(textgrid, filename, format, encoding, **kwargs)
390 with codecs.open(filename, 'w', encoding) as f:
391 if format in _EXPORT_FORMATS:
--> 392 f.write(_EXPORT_FORMATS[format](textgrid, **kwargs))
393 else:
394 raise Exception('Unknown output format: {0}'.format(format))

/usr/local/lib/python2.7/site-packages/tgt/io.pyc in export_to_short_textgrid(textgrid)
241 textgrid_corrected = correct_start_end_times_and_fill_gaps(textgrid)
242 for tier in textgrid_corrected:
--> 243
result += ['"' + tier.tier_type() + '"',

244                    '"' + escape_text(tier.name) + '"',
245
                tier.start_time, tier.end_time, len(tier)]

AttributeError: 'Interval' object has no attribute 'tier_type'

RESPONSE:

the problem you encounter is solved easily. To add the shifted tier, you did the following:

tg.add_tiers(tierShifted)

This method, however, expects a list of tiers, not a single tier. You have to do the following instead

tg.add_tier(tierShifted)

or

tg.add_tiers([tierShifted])