Sphinx 4 - 0.00% accuracy using hub 4.

[Uthman]

Hello

I'm trying to run a test using hub 4 but I'm constantly getting 0% accurracy. The summary also diplays there has been 4 errors but with giving details of them. I've followed the instructions on the web site but still no luck. Nothing is being displayed for the hyp. Here is the out I'm getting.

Buildfile: build.xml

hub4_trigram:
[java] 03:47.741 INFO logMath Log base is 1.0001
[java] 03:47.756 INFO logMath Using AddTable when adding logs
[java] 03:47.756 INFO logMath LogAdd table has 99022 entries.
[java] 03:47.897 INFO dictionary Loading dictionary from: file:/C:/sphinx4/bld/classes/edu/cmu/sphinx/model/acoustic/HUB4_8gau_13dCep_16k_40mel_133Hz_6855Hz/cmudict.06d
[java] 03:47.334 INFO dictionary Loading filler dictionary from: file:/C:/sphinx4/bld/classes/edu/cmu/sphinx/model/acoustic/HUB4_8gau_13dCep_16k_40mel_133Hz_6855Hz/fillerdict
[java] 03:47.631 INFO unitManager CI Unit: +BREATH+
[java] 03:47.631 INFO unitManager CI Unit: +COUGH+
[java] 03:47.631 INFO unitManager CI Unit: +GARBAGE+
[java] 03:47.631 INFO unitManager CI Unit: +NOISE+
[java] 03:47.631 INFO unitManager CI Unit: +SMACK+
[java] 03:47.631 INFO unitManager CI Unit: +UH+
[java] 03:47.631 INFO unitManager CI Unit: +UHUM+
[java] 03:47.631 INFO unitManager CI Unit: +UM+
[java] 03:47.631 INFO unitManager CI Unit: AA
[java] 03:47.631 INFO unitManager CI Unit: AE
[java] 03:47.631 INFO unitManager CI Unit: AH
[java] 03:47.631 INFO unitManager CI Unit: AO
[java] 03:47.631 INFO unitManager CI Unit: AW
[java] 03:47.631 INFO unitManager CI Unit: AY
[java] 03:47.631 INFO unitManager CI Unit: B
[java] 03:47.631 INFO unitManager CI Unit: CH
[java] 03:47.631 INFO unitManager CI Unit: D
[java] 03:47.647 INFO unitManager CI Unit: DH
[java] 03:47.647 INFO unitManager CI Unit: EH
[java] 03:47.647 INFO unitManager CI Unit: ER
[java] 03:47.647 INFO unitManager CI Unit: EY
[java] 03:47.647 INFO unitManager CI Unit: F
[java] 03:47.647 INFO unitManager CI Unit: G
[java] 03:47.647 INFO unitManager CI Unit: HH
[java] 03:47.647 INFO unitManager CI Unit: IH
[java] 03:47.647 INFO unitManager CI Unit: IY
[java] 03:47.647 INFO unitManager CI Unit: JH
[java] 03:47.647 INFO unitManager CI Unit: K
[java] 03:47.647 INFO unitManager CI Unit: L
[java] 03:47.647 INFO unitManager CI Unit: M
[java] 03:47.647 INFO unitManager CI Unit: N
[java] 03:47.647 INFO unitManager CI Unit: NG
[java] 03:47.647 INFO unitManager CI Unit: OW
[java] 03:47.647 INFO unitManager CI Unit: OY
[java] 03:47.647 INFO unitManager CI Unit: P
[java] 03:47.647 INFO unitManager CI Unit: R
[java] 03:47.647 INFO unitManager CI Unit: S
[java] 03:47.647 INFO unitManager CI Unit: SH
[java] 03:47.647 INFO unitManager CI Unit: T
[java] 03:47.647 INFO unitManager CI Unit: TH
[java] 03:47.647 INFO unitManager CI Unit: UH
[java] 03:47.647 INFO unitManager CI Unit: UW
[java] 03:47.647 INFO unitManager CI Unit: V
[java] 03:47.647 INFO unitManager CI Unit: W
[java] 03:47.647 INFO unitManager CI Unit: Y
[java] 03:47.647 INFO unitManager CI Unit: Z
[java] 03:47.647 INFO unitManager CI Unit: ZH
[java] 03:47.553 INFO sphinx3Loader ModelLoader
[java] 03:47.553 INFO sphinx3Loader Pool 6000senones/hub4opensrc.cd_continuous_8gau/means Entries: 49152
[java] 03:47.553 INFO sphinx3Loader Pool 6000senones/hub4opensrc.cd_continuous_8gau/variances Entries: 49152
[java] 03:47.553 INFO sphinx3Loader Pool 6000senones/hub4opensrc.cd_continuous_8gau/transition_matrices Entries: 48
[java] 03:47.553 INFO sphinx3Loader Pool senones Entries: 6144
[java] 03:47.553 INFO sphinx3Loader Pool meanTransformationMatrix Entries: 1
[java] 03:47.553 INFO sphinx3Loader Pool meanTransformationMatrix Entries: 1
[java] 03:47.553 INFO sphinx3Loader Pool varianceTransformationMatrix Entries: 1
[java] 03:47.553 INFO sphinx3Loader Pool varianceTransformationMatrix Entries: 1
[java] 03:47.553 INFO sphinx3Loader Pool 6000senones/hub4opensrc.cd_continuous_8gau/mixture_weights Entries: 6144
[java] 03:47.553 INFO sphinx3Loader Pool senones Entries: 6144
[java] 03:47.553 INFO sphinx3Loader Context Independent Unit Entries: 48
[java] 03:47.553 INFO sphinx3Loader HMM Manager: 133548 hmms
[java] 03:47.553 INFO hub4 CompositeSenoneSequences: 0.

I than get alot of warning of missing words.
[java] 03:47.818 INFO trigramModel Missing word: abidjan
[java] 03:47.818 WARNING dictionary Missing word: abimael
[java] in
edu.cmu.sphinx.linguist.dictionary.FastDictionary:getWord-dictionary

[java] 03:47.114 INFO threadedScorer # of scoring threads: 2
[java] # ----------------------------- Timers----------------------------------------
[java] # Name Count CurTime MinTime MaxTime AvgTime TotTime
[java] DictionaryLoad 1 0.4370s 0.4370s 0.4370s 0.4370s 0.4370s
[java] AM_Load 1 5.2190s 5.2190s 5.2190s 5.2190s 5.2190s
[java] LM Load 1 1.3740s 1.3740s 1.3740s 1.3740s 1.3740s
[java] compile 1 7.1560s 7.1560s 7.1560s 7.1560s 7.1560s
[java] buildHmmPool 1 1.7030s 1.7030s 1.7030s 1.7030s 1.7030s
[java] Create HMMTree 1 5.4060s 5.4060s 5.4060s 5.4060s 5.4060s
[java] ============ config =============
[java] batch:
[java] inputDataProcessors = streamDataSource
[java] recognizer = wordRecognizer
[java] whichBatch = [DEFAULT]
[java] count = [DEFAULT]
[java] usePooledBatchManager = [DEFAULT]
[java] logLevel = [DEFAULT]
[java] skip = 0
[java] totalBatches = [DEFAULT]
[java] wordRecognizer:
[java] monitors = accuracyTracker, speedTracker, memoryTracker, recognizerMonitor, beamFinder
[java] logLevel = [DEFAULT]
[java] decoder = decoder
[java] decoder:
[java] logLevel = [DEFAULT]
[java] featureBlockSize = [DEFAULT]
[java] searchManager = wordPruningSearchManager
[java] wordPruningSearchManager:
[java] growSkipInterval = 8
[java] pruner = trivialPruner
[java] acousticLookaheadFrames = 2.0
[java] linguist = lexTreeLinguist
[java] showTokenCount = [DEFAULT]
[java] buildWordLattice = false
[java] maxLatticeEdges = [DEFAULT]
[java] relativeBeamWidth = 1E-60
[java] keepAllTokens = true
[java] logLevel = [DEFAULT]
[java] scorer = threadedScorer
[java] activeListManager = activeListManager
[java] checkStateOrder = false
[java] logMath = logMath
[java] logMath:
[java] logLevel = [DEFAULT]
[java] useAddTable = true
[java] logBase = 1.0001
[java] lexTreeLinguist:
[java] wantUnigramSmear = true
[java] unitManager = unitManager
[java] generateUnitStates = false
[java] acousticModel = hub4
[java] dictionary = dictionary
[java] wordInsertionProbability = .2
[java] languageWeight = 11
[java] fillerInsertionProbability = .02
[java] logLevel = [DEFAULT]
[java] languageModel = trigramModel
[java] fullWordHistories = [DEFAULT]
[java] unitInsertionProbability = [DEFAULT]
[java] silenceInsertionProbability = .1
[java] logMath = logMath
[java] addFillerWords = true
[java] unigramSmearWeight = 1
[java] cacheSize = 0
[java] hub4:
[java] useComposites = [DEFAULT]
[java] unitManager = unitManager
[java] logLevel = [DEFAULT]
[java] loader = sphinx3Loader
[java] sphinx3Loader:
[java] unitManager = unitManager
[java] propertiesFile = [DEFAULT]
[java] MixtureComponentScoreFloor = [DEFAULT]
[java] sparseForm = [DEFAULT]
[java] isBinary = [DEFAULT]
[java] logLevel = [DEFAULT]
[java] vectorLength = [DEFAULT]
[java] mixtureWeightFloor = [DEFAULT]
[java] dataLocation = [DEFAULT]
[java] useCDUnits = [DEFAULT]
[java] logMath = logMath
[java] varianceFloor = [DEFAULT]
[java] modelDefinition = [DEFAULT]
[java] unitManager:
[java] logLevel = [DEFAULT]
[java] trigramModel:
[java] clearCachesAfterUtterance = [DEFAULT]
[java] location = C:/sphinx4/tests/performance/hub4/language_model.arpaformat.DMP
[java] applyLanguageWeightAndWip = [DEFAULT]
[java] format = [DEFAULT]
[java] dictionary = dictionary
[java] languageWeight = [DEFAULT]
[java] wordInsertionProbability = [DEFAULT]
[java] unigramWeight = .5
[java] bigramCacheSize = [DEFAULT]
[java] trigramCacheSize = [DEFAULT]
[java] maxDepth = 3
[java] logLevel = [DEFAULT]
[java] fullSmear = [DEFAULT]
[java] logMath = logMath
[java] queryLogFile = [DEFAULT]
[java] dictionary:
[java] unitManager = unitManager
[java] dictionaryPath = resource:/edu.cmu.sphinx.model.acoustic.HUB4_8gau_13dCep_16k_40mel_133Hz_6855Hz.Model!/edu/cmu/sphinx/model/acoustic/HUB4_8gau_13dCep_16k_40mel_133Hz_6855Hz/cmudict.06d
[java] wordReplacement = [DEFAULT]
[java] logLevel = [DEFAULT]
[java] allowMissingWords = false
[java] fillerPath = resource:/edu.cmu.sphinx.model.acoustic.HUB4_8gau_13dCep_16k_40mel_133Hz_6855Hz.Model!/edu/cmu/sphinx/model/acoustic/HUB4_8gau_13dCep_16k_40mel_133Hz_6855Hz/fillerdict
[java] addSilEndingPronunciation = false
[java] createMissingWords = [DEFAULT]
[java] trivialPruner:
[java] logLevel = [DEFAULT]
[java] threadedScorer:
[java] logLevel = [DEFAULT]
[java] numThreads = 0
[java] scoreablesKeepFeature = true
[java] minScoreablesPerThread = 10
[java] frontend = mfcFrontEnd
[java] acousticGain = [DEFAULT]
[java] isCpuRelative = true
[java] mfcFrontEnd:
[java] logLevel = [DEFAULT]
[java] pipeline = streamDataSource, premphasizer, windower, fft, melFilterBank, dct, batchCMN, featureExtraction
[java] streamDataSource:
[java] sampleRate = 16000
[java] bigEndianData = [DEFAULT]
[java] logLevel = [DEFAULT]
[java] signedData = [DEFAULT]
[java] bitsPerSample = [DEFAULT]
[java] bytesPerRead = [DEFAULT]
[java] premphasizer:
[java] logLevel = [DEFAULT]
[java] factor = [DEFAULT]
[java] windower:
[java] logLevel = [DEFAULT]
[java] alpha = [DEFAULT]
[java] windowSizeInMs = [DEFAULT]
[java] windowShiftInMs = [DEFAULT]
[java] fft:
[java] logLevel = [DEFAULT]
[java] numberFftPoints = [DEFAULT]
[java] melFilterBank:
[java] numberFilters = [DEFAULT]
[java] minimumFrequency = [DEFAULT]
[java] logLevel = [DEFAULT]
[java] maximumFrequency = [DEFAULT]
[java] dct:
[java] numberFilters = [DEFAULT]
[java] cepstrumLength = [DEFAULT]
[java] logLevel = [DEFAULT]
[java] batchCMN:
[java] logLevel = [DEFAULT]
[java] featureExtraction:
[java] logLevel = [DEFAULT]
[java] windowSize = [DEFAULT]
[java] activeListManager:
[java] logLevel = [DEFAULT]
[java] activeListFactories = unitExitActiveList, wordActiveList, wordActiveList, activeList, activeList, activeList
[java] checkPriorListsEmpty = [DEFAULT]
[java] unitExitActiveList:
[java] relativeBeamWidth = 1E-60
[java] logLevel = [DEFAULT]
[java] logMath = logMath
[java] absoluteBeamWidth = -1
[java] wordActiveList:
[java] relativeBeamWidth = 1E-25
[java] logLevel = [DEFAULT]
[java] maxPathsPerWord = [DEFAULT]
[java] logMath = logMath
[java] absoluteBeamWidth = 21
[java] maxFillerWords = [DEFAULT]
[java] activeList:
[java] relativeBeamWidth = 1E-60
[java] logLevel = [DEFAULT]
[java] logMath = logMath
[java] absoluteBeamWidth = 25000
[java] accuracyTracker:
[java] recognizer = wordRecognizer
[java] showSummary = [DEFAULT]
[java] logLevel = [DEFAULT]
[java] showFullPath = [DEFAULT]
[java] showAlignedResults = false
[java] showDetails = [DEFAULT]
[java] showRawResults = false
[java] showResults = [DEFAULT]
[java] speedTracker:
[java] recognizer = wordRecognizer
[java] showSummary = [DEFAULT]
[java] logLevel = [DEFAULT]
[java] showTimers = true
[java] showDetails = [DEFAULT]
[java] showResponseTime = [DEFAULT]
[java] frontend = mfcFrontEnd
[java] memoryTracker:
[java] recognizer = wordRecognizer
[java] showSummary = [DEFAULT]
[java] logLevel = [DEFAULT]
[java] showDetails = [DEFAULT]
[java] recognizerMonitor:
[java] recognizer = wordRecognizer
[java] logLevel = [DEFAULT]
[java] deallocatedMonitors = [DEFAULT]
[java] allocatedMonitors = configMonitor
[java] configMonitor:
[java] showConfigAsGDL = [DEFAULT]
[java] logLevel = [DEFAULT]
[java] showConfig = true
[java] saveConfigAsXML = true
[java] showConfigAsHTML = [DEFAULT]
[java] beamFinder:
[java] recognizer = wordRecognizer
[java] showSummary = true
[java] logLevel = [DEFAULT]
[java] showDetails = true
[java] enable = false
[java] logMath = logMath
[java] 03:47.223 INFO batch BatchDecoder: decoding files in C:/sphinx4/tests/performance/hub4/test.batch
[java] 03:47.333 INFO batch WAVE (.wav) file, byte length: 80058, data format: PCM_SIGNED 16000.0 Hz, 16 bit, mono, 2 bytes/frame, little-endian, frame length: 40000
[java]
[java] REF: this is a test
[java] HYP:
[java]
[java] Accuracy: 0.000% Errors: 4 (Sub: 0 Ins: 0 Del: 4)
[java] Words: 4 Matches: 0 WER: 100.000%
[java] Sentences: 1 Matches: 0 SentenceAcc: 0.000%
[java] This Time Audio: 2.50s Proc: 4.42s Speed: 1.77 X real time
[java] Total Time Audio: 2.50s Proc: 4.42s Speed: 1.77 X real time
[java] Mem Total: 197.81 Mb Free: 39.27 Mb
[java] Used: This: 158.55 Mb Avg: 158.55 Mb Max: 158.55 Mb

Can any one help me to see whats wrong?

[Nickolay V. Shmyrev]

Your configuration is wrong. What are you trying to do?

The question of LVCSR with sphinx4 was discussed many times on this forum with examples. I suggest you to look there first. Also I'd recommend you to avoid duplicated posting on different forums.

[Uthman]

Thank you for replying and sorry for the duplicated post.

I'm trying to run a regression test on hub 4 to see how it performs? I followed the instructions here http://cmusphinx.sourceforge.net/sphinx4/#large_vocab_test but don't see be having any luck.

I'm trying to see how hub 4 performs. I'm trying to decide whether to move to hub 4 or stick with WSJ (WSJ works for me but has very poor accuracy).

[Uthman]

I've managed to modify the transcriber demo to use both hub4 and WSJ. I'm getting very poor results for both of these.

These are the types of results I'm getting:

File 1
Audio Properties:WAVE (.wav) file, byte length: 80058, data format: PCM_SIGNED 16000.0 Hz, 16 bit, mono, 2 bytes/frame, little-endian, frame length: 40000
Audio File Transcript: "This is a test".
Result: "this suzette istook"

File 2
Audio Properties:WAVE (.wav) file, byte length: 216058, data format: PCM_SIGNED 16000.0 Hz, 16 bit, mono, 2 bytes/frame, little-endian, frame length: 108000
Audio File Transcript:"Near a slate mine in north wales theres an area where the red dragon fought the white dragon and won".
Result: "here's langbein your wales' isn't urea whether ejecting forgoing tracking one"

These are the results when using hub4.

Obvisoulsy the actual results are not very good. Can you see where I'm going wrong. The results uing WSJ are just as bad?

test.wav http://www.mediafire.com/?ez4exjndhxl

dragon1.wav http://www.mediafire.com/?zimyrsilf9s

hub4.config.xml http://www.mediafire.com/?vdbjzdwgsnf

Transcriber.java http://www.mediafire.com/?wnxhdrowtdn

[Nickolay V. Shmyrev]

You need bigger word insertion probability (around 0.7 for wsj). Check the exact values in hub4 config in tests/regressions. It should run quite well if you'll reproduce the situation.

Also remember that hub4 is a quite specific task with a specific language model. I suggest you to look on forum examples instead first.