I finally was able to run a modified version of the transcriber demo in order to use HUB4 model for transcription.
Now I have some English newscast programmes and I'm trying to transcribe them, but results are very poor: large parts of speech seem not to be transcribed and the transcribed part is not so good.
Do you think that there are parameters that can be tuned in order to have better performances or the problems is linked to the model?
I'm not sure how do you transcribe newscast since your config works with microphone. If you are decoding in a batch mode take care about sampling rate first of all. If everything is setup correctly first of all you have to train a language model for your domain.
P.S. Please don't start a new topic for every question.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Thanks for your help.
There were some useless parts in the configuration file (the microphone part was there but was not used) but the frontend uses a streamDataSource.
I've changed from LiveCMN to BatchCMN and the results are now a little bit better. Do you think that changing parameters (absoluteBeanWidth, absoluteWordBeanWidth and so on) can improve performances? I've seen that speech is not correctly recognized when the voice is not completely clear or there are noises in the background. I think this problem is due to the acoustic model used. I mean, probably the HUB4 acoustic model was created starting from audio recorded by speakers reading sentences. Have you any idea about which materials where used for creating the HUB4 acoustic model? Probably I have to create my own acoustic model that is nearest to the context in which I want to use it (newscast programmes).
Unfortunatly I'm not able to join irc channels at work (security reasons).
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I suggest you to prepare a small demo of your work and upload it somewhere. We'll look and suggest you something. Most probably you'll have to use your own language model and probably adaptation. But you could make much more silly mistake.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi,
i thnk the recognition speed is slow when absoluteBeamWidth value="-1". -1(mean infiniti in sphinx4 config term)is for a small task like digit or alphabet or simple command. For big task, the value should around 1000-5000.
For the acoustic model i dn't have any idea ..am using the HUB4 itself..can u please tell me how much memory you allocate to your sphinx4 application??Since my pc memory is very low it's taking lotz of time to decode and mostly due to hub4 size!!.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
I'm trying to convert short podcasts with different speakers from 16k mono 256kbps wav to text, but the accuracy is very poor.
Increasing the beamwidth values doesn't seem to improve the accuracy much.
Following the steps of another forum post, and since the speakers are random/unknown, I modified the Transcriber demo config file to use the Wall Street Journal language model (wsj5k.DMP).
If there's a better solution or there's anything obviously wrong with the config file posted below, please let me know.
Hello,
The recognition is not done live..thnk you should use BatchCMN instead of LiveCMN.Am not sure abt this?can any1 please clear this out??what's the difference between these 2 and what if we use Batch instead of Live??
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
These are the parameters I'm using for my application: -Xmx1024M -Xms512M With these parameters I can transcribe a wav file of 27 minutes in 2 hours and a half.
I have to find an host where I can place the demo.
When you speak of adaptation do you mean adaptation of the acoustic model after the train? I have read that something like this is done by some commercial software of speech recognition.
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
Hi,
I finally was able to run a modified version of the transcriber demo in order to use HUB4 model for transcription.
Now I have some English newscast programmes and I'm trying to transcribe them, but results are very poor: large parts of speech seem not to be transcribed and the transcribed part is not so good.
Do you think that there are parameters that can be tuned in order to have better performances or the problems is linked to the model?
Here is my configuration file:
<config>
<!-- ******** -->
<!-- frequently tuned properties -->
<!-- ******** -->
<property name="absoluteBeamWidth" value="5000"/>
<property name="relativeBeamWidth" value="1E-80"/>
<property name="absoluteWordBeamWidth" value="80"/>
<property name="relativeWordBeamWidth" value="1E-60"/>
<property name="wordInsertionProbability" value="1E-16"/>
<property name="languageWeight" value="7.0"/>
<property name="silenceInsertionProbability" value=".1"/>
<property name="frontend" value="epFrontEnd"/>
<property name="recognizer" value="recognizer"/>
<property name="showCreations" value="false"/>
</config>
I'm not sure how do you transcribe newscast since your config works with microphone. If you are decoding in a batch mode take care about sampling rate first of all. If everything is setup correctly first of all you have to train a language model for your domain.
P.S. Please don't start a new topic for every question.
if you need quick help for such simple questions you can join #cmusphinx irc channel on freenode.
Thanks for your help.
There were some useless parts in the configuration file (the microphone part was there but was not used) but the frontend uses a streamDataSource.
I've changed from LiveCMN to BatchCMN and the results are now a little bit better. Do you think that changing parameters (absoluteBeanWidth, absoluteWordBeanWidth and so on) can improve performances? I've seen that speech is not correctly recognized when the voice is not completely clear or there are noises in the background. I think this problem is due to the acoustic model used. I mean, probably the HUB4 acoustic model was created starting from audio recorded by speakers reading sentences. Have you any idea about which materials where used for creating the HUB4 acoustic model? Probably I have to create my own acoustic model that is nearest to the context in which I want to use it (newscast programmes).
Unfortunatly I'm not able to join irc channels at work (security reasons).
Hub4 is fine as well as wsj
I suggest you to prepare a small demo of your work and upload it somewhere. We'll look and suggest you something. Most probably you'll have to use your own language model and probably adaptation. But you could make much more silly mistake.
Hi,
i thnk the recognition speed is slow when absoluteBeamWidth value="-1". -1(mean infiniti in sphinx4 config term)is for a small task like digit or alphabet or simple command. For big task, the value should around 1000-5000.
For the acoustic model i dn't have any idea ..am using the HUB4 itself..can u please tell me how much memory you allocate to your sphinx4 application??Since my pc memory is very low it's taking lotz of time to decode and mostly due to hub4 size!!.
I'm experiencing essentially the same problem.
I'm trying to convert short podcasts with different speakers from 16k mono 256kbps wav to text, but the accuracy is very poor.
Increasing the beamwidth values doesn't seem to improve the accuracy much.
Following the steps of another forum post, and since the speakers are random/unknown, I modified the Transcriber demo config file to use the Wall Street Journal language model (wsj5k.DMP).
If there's a better solution or there's anything obviously wrong with the config file posted below, please let me know.
<?xml version="1.0" encoding="UTF-8"?>
<!--
Sphinx-4 Configuration file
-->
<!-- ******** -->
<!-- an4 configuration file -->
<!-- ******** -->
<config>
<!--*** -->
<!-- frequently tuned properties -->
<!-- ******** -->
<property name="logLevel" value="OFF"/>
<!-- <property name="relativeBeamWidth" value="1E-10" /> -->
<property name="absoluteWordBeamWidth" value="5000" />
<property name="relativeWordBeamWidth" value="1E-200" />
<property name="wordInsertionProbability" value="1E-36" />
<property name="languageWeight" value="8" />
<property name="silenceInsertionProbability" value=".3" />
<property name="acousticLookahead" value="1.7" />
<property name="absoluteBeamWidth" value="5000" />
<property name="relativeBeamWidth" value="1E-230" />
<property name="frontend" value="epFrontEnd"/>
<property name="recognizer" value="recognizer"/>
<property name="showCreations" value="false"/>
<!-- ******** -->
<!-- word recognizer configuration -->
<!-- ******** -->
<component name="recognizer"
type="edu.cmu.sphinx.recognizer.Recognizer">
<property name="decoder" value="decoder"/>
<propertylist name="monitors">
<item>accuracyTracker </item>
<item>speedTracker </item>
<item>memoryTracker </item>
</propertylist>
</component>
<component name="lexTreeLinguist"
type="edu.cmu.sphinx.linguist.lextree.LexTreeLinguist">
<property name="silenceInsertionProbability"
value="${silenceInsertionProbability}" />
<property name="wantUnigramSmear" value="true" />
<property name="fillerInsertionProbability" value=".02" />
<property name="addFillerWords" value="true" />
<property name="acousticModel" value="wsj" />
<property name="languageModel" value="trigramModel" />
<property name="wordInsertionProbability"
value="${wordInsertionProbability}" />
<property name="languageWeight" value="14" />
<property name="logMath" value="logMath" />
<property name="dictionary" value="dictionary" />
<property name="unigramSmearWeight" value="1" />
<property name="cacheSize" value="0" />
<property name="generateUnitStates" value="false" />
<property name="unitManager" value="unitManager" />
</component>
<!-- ******** -->
<!-- The Decoder configuration -->
<!-- ******** -->
<component name="decoder" type="edu.cmu.sphinx.decoder.Decoder">
<property name="searchManager" value="wordPruningSearchManager"/>
</component>
-->
-->
-->
type="edu.cmu.sphinx.decoder.search.WordPruningBreadthFirstSearchManager">
<property name="scorer" value="threadedScorer" />
<property name="pruner" value="trivialPruner" />
<property name="acousticLookaheadFrames" value="2.0" />
<property name="logMath" value="logMath" />
<property name="activeListManager" value="activeListManager" />
<property name="buildWordLattice" value="true" />
<property name="maxLatticeEdges" value ="50" />
<property name="relativeBeamWidth" value="1E-60" />
<property name="growSkipInterval" value="8" />
<property name="linguist" value="lexTreeLinguist" />
<property name="checkStateOrder" value="false" />
<property name="keepAllTokens" value="true" />
</component>
<component name="activeList"
type="edu.cmu.sphinx.decoder.search.PartitionActiveListFactory">
<property name="logMath" value="logMath"/>
<property name="absoluteBeamWidth" value="${absoluteBeamWidth}"/>
<property name="relativeBeamWidth" value="${relativeBeamWidth}"/>
</component>
<component name="trivialPruner"
type="edu.cmu.sphinx.decoder.pruner.SimplePruner"/>
<component name="threadedScorer"
type="edu.cmu.sphinx.decoder.scorer.ThreadedAcousticScorer">
<property name="frontend" value="${frontend}"/>
<property name="isCpuRelative" value="true"/>
<property name="numThreads" value="0"/>
<property name="minScoreablesPerThread" value="10"/>
<property name="scoreablesKeepFeature" value="true"/>
</component>
<!-- ******** -->
<!-- acoustic model -->
<!-- ******** -->
<component name="wsj"
type="edu.cmu.sphinx.model.acoustic.WSJ_8gau_13dCep_16k_40mel_130Hz_6800Hz.Model">
<property name="loader" value="wsjLoader" />
<property name="unitManager" value="unitManager" />
</component>
<!-- ******** -->
<!-- sphinx3Loader -->
<!-- ******** -->
<component name="wsjLoader"
type="edu.cmu.sphinx.model.acoustic.WSJ_8gau_13dCep_16k_40mel_130Hz_6800Hz.ModelLoader">
<property name="logMath" value="logMath" />
<property name="unitManager" value="unitManager" />
</component>
<!-- ******** -->
<!-- trigramModel -->
<!-- ******** -->
<component name="trigramModel"
type="edu.cmu.sphinx.linguist.language.ngram.large.LargeTrigramModel">
<property name="unigramWeight" value=".5" />
<property name="maxDepth" value="3" />
<property name="logMath" value="logMath" />
<property name="dictionary" value="dictionary" />
<property name="location" value="/usr/src/wsj5k.DMP" />
</component>
<!-- ******** -->
<!-- The Dictionary configuration -->
<!-- ******** -->
<component name="dictionary"
type="edu.cmu.sphinx.linguist.dictionary.FastDictionary">
<property name="dictionaryPath"
value="resource:/edu.cmu.sphinx.model.acoustic.WSJ_8gau_13dCep_16k_40mel_130Hz_6800Hz.Model!/edu/cmu/sphinx/model/acoustic/WSJ_8gau_13dCep_16k_40mel_130Hz_6800Hz/dict/cmudict.0.6d"/>
<property name="fillerPath"
value="resource:/edu.cmu.sphinx.model.acoustic.WSJ_8gau_13dCep_16k_40mel_130Hz_6800Hz.Model!/edu/cmu/sphinx/model/acoustic/WSJ_8gau_13dCep_16k_40mel_130Hz_6800Hz/dict/fillerdict"/>
<property name="addSilEndingPronunciation" value="false"/>
<property name="wordReplacement" value="<sil>"/>
<property name="allowMissingWords" value="false"/>
<property name="unitManager" value="unitManager"/>
</component>
<!-- ******** -->
<!-- The unit manager configuration -->
<!-- ******** -->
<component name="unitManager"
type="edu.cmu.sphinx.linguist.acoustic.UnitManager"/>
<!-- ******** -->
<!-- The live frontend configuration -->
<!-- ******** -->
<component name="epFrontEnd" type="edu.cmu.sphinx.frontend.FrontEnd">
<propertylist name="pipeline">
<item>streamDataSource </item>
<item>speechClassifier </item>
<item>speechMarker </item>
<item>nonSpeechDataFilter </item>
<item>premphasizer </item>
<item>windower </item>
<item>fft </item>
<item>melFilterBank </item>
<item>dct </item>
<item>liveCMN </item>
<item>featureExtraction </item>
</propertylist>
</component>
<!-- ******** -->
<!-- The frontend pipelines -->
<!-- ******** -->
<component name="streamDataSource"
type="edu.cmu.sphinx.frontend.util.StreamDataSource">
<property name="sampleRate" value="16000"/>
<property name="bitsPerSample" value="16"/>
<property name="bigEndianData" value="false"/>
<property name="signedData" value="true"/>
<property name="bytesPerRead" value="320"/>
</component>
<component name="speechClassifier"
type="edu.cmu.sphinx.frontend.endpoint.SpeechClassifier">
<property name="threshold" value="13"/>
</component>
<component name="nonSpeechDataFilter"
type="edu.cmu.sphinx.frontend.endpoint.NonSpeechDataFilter"/>
<component name="speechMarker"
type="edu.cmu.sphinx.frontend.endpoint.SpeechMarker" >
<property name="speechTrailer" value="50"/>
</component>
<component name="premphasizer"
type="edu.cmu.sphinx.frontend.filter.Preemphasizer"/>
<component name="windower"
type="edu.cmu.sphinx.frontend.window.RaisedCosineWindower">
</component>
<component name="fft"
type="edu.cmu.sphinx.frontend.transform.DiscreteFourierTransform">
</component>
<component name="melFilterBank"
type="edu.cmu.sphinx.frontend.frequencywarp.MelFrequencyFilterBank">
</component>
<component name="dct"
type="edu.cmu.sphinx.frontend.transform.DiscreteCosineTransform"/>
<component name="liveCMN"
type="edu.cmu.sphinx.frontend.feature.LiveCMN"/>
<component name="featureExtraction"
type="edu.cmu.sphinx.frontend.feature.DeltasFeatureExtractor"/>
<!-- ********* -->
<!-- monitors -->
<!-- ********* -->
<component name="accuracyTracker"
type="edu.cmu.sphinx.instrumentation.AccuracyTracker">
<property name="recognizer" value="${recognizer}"/>
<property name="showAlignedResults" value="false"/>
<property name="showRawResults" value="false"/>
</component>
<component name="memoryTracker"
type="edu.cmu.sphinx.instrumentation.MemoryTracker">
<property name="recognizer" value="${recognizer}"/>
<property name="showSummary" value="false"/>
<property name="showDetails" value="false"/>
</component>
<component name="speedTracker"
type="edu.cmu.sphinx.instrumentation.SpeedTracker">
<property name="recognizer" value="${recognizer}"/>
<property name="frontend" value="${frontend}"/>
<property name="showSummary" value="true"/>
<property name="showDetails" value="false"/>
</component>
-->
-->
-->
type="edu.cmu.sphinx.decoder.search.SimpleActiveListManager">
<item>unitExitActiveList</item>
<item>wordActiveList</item>
<item>wordActiveList</item>
<item>activeList</item>
<item>activeList</item>
<item>activeList</item>
</propertylist>
</component>
<!-- ********
-->
-->
-->
type="edu.cmu.sphinx.decoder.search.PartitionActiveListFactory">
<property name="absoluteBeamWidth" value="-1" />
<property name="logMath" value="logMath" />
<property name="relativeBeamWidth" value="${relativeBeamWidth}" />
</component>
-->
-->
-->
type="edu.cmu.sphinx.decoder.search.WordActiveListFactory">
<property name="absoluteBeamWidth" value="21" />
<property name="logMath" value="logMath" />
<property name="relativeBeamWidth" value="1E-25" />
</component>
-->
-->
-->
type="edu.cmu.sphinx.instrumentation.RecognizerMonitor">
<property name="recognizer" value="${recognizer}" />
<item>configMonitor</item>
</propertylist>
</component>
<!-- ********* -->
<!-- Miscellaneous components -->
<!-- ********* -->
<component name="logMath" type="edu.cmu.sphinx.util.LogMath">
<property name="logBase" value="1.0001"/>
<property name="useAddTable" value="true"/>
</component>
<component name="confidenceScorer"
type="edu.cmu.sphinx.result.MAPConfidenceScorer">
<property name="languageWeight" value="${languageWeight}"/>
</component>
</config>
> I modified the Transcriber demo config file to use the Wall Street Journal language model (wsj5k.DMP).
This is obviously wrong.
Hi,
Could you please elaborate?
I followed the config in the first post of this thread, which you said looks otherwise ok.
http://sourceforge.net/forum/forum.php?thread_id=1935440&forum_id=5471
Hello,
The recognition is not done live..thnk you should use BatchCMN instead of LiveCMN.Am not sure abt this?can any1 please clear this out??what's the difference between these 2 and what if we use Batch instead of Live??
These are the parameters I'm using for my application: -Xmx1024M -Xms512M With these parameters I can transcribe a wav file of 27 minutes in 2 hours and a half.
I have to find an host where I can place the demo.
When you speak of adaptation do you mean adaptation of the acoustic model after the train? I have read that something like this is done by some commercial software of speech recognition.
> I can transcribe a wav file of 27 minutes in 2 hours and a half.
Beams could be smaller, it will work faster
> I have to find an host where I can place the demo.
Use mediafire.com for example. It shouldn't be two hours, just a few minutes of speech
> I have read that something like this is done by some commercial software of speech recognition.
Not only commercial software :) Sphinx4 doesn't support mllr but sphinx3 does:
http://www.speech.cs.cmu.edu/cmusphinx/moinmoin/AcousticModelAdaptation