[{"data":1,"prerenderedAt":209},["ShallowReactive",2],{"\u002Fsphinx-notes":3},{"id":4,"title":5,"body":6,"createdAt":198,"description":12,"draft":199,"extension":200,"image":201,"mastodonThread":201,"meta":202,"navigation":203,"path":204,"seo":205,"stem":206,"tags":207,"updatedAt":198,"__hash__":208},"content\u002FSphinx Notes.md","Sphinx_Notes",{"type":7,"value":8,"toc":192},"minimark",[9,13,18,21,24,27,31,41,48,59,74,95,110,128,143,171,175],[10,11,12],"p",{},"These are just some collection of notes about getting Sphinx to do what I want it to do. Wish me luck.",[14,15,17],"h2",{"id":16},"_20041006","2004.10.06",[10,19,20],{},"Working on using the hub4 regression test to make this thing work. I converted a voicemail to a raw audio file with",[10,22,23],{},"sox bethmom-packet8_voiceMessage.au -r 16000 -w -s -x bethmom2.raw",[10,25,26],{},"the trickiest part to figure out was the -x to switch endianness. I'm running through another example now to see how well it works, then I'll run the above through. It is certainly not fast the way I have it set up now.",[14,28,30],{"id":29},"component-overview","Component overview",[10,32,33,34,40],{},"(From ",[35,36,37],"a",{"href":37,"rel":38},"http:\u002F\u002Fharvee.org\u002Fpipermail\u002Fossri\u002F2004-September\u002F000826.html",[39],"nofollow",")",[10,42,43,44,47],{},"There are a number of components that play a role here.",[45,46],"br",{},"\nIn practice, we care about the following:",[49,50,55],"pre",{"className":51,"code":53,"language":54},[52],"language-text","1) The acoustic model and phone set\n2) The phone set\n3) The pronunciation lexicon\n4) The language model\n","text",[56,57,53],"code",{"__ignoreMap":58},"",[10,60,61,62,64,65,67,68,70,71,73],{},"The phone set defines the sub-word units that will be modeled",[45,63],{},"\nby the acoustic model.  These sub-word units can be things",[45,66],{},"\nsuch as phonemes (e.g., \"ah\", \"ch,\" \"k\", ...), word dependent",[45,69],{},"\nphonemes (e.g., the \"t\" in the word \"eight\"), syllables, and",[45,72],{},"\nso on.",[10,75,76,77,79,80,82,83,85,86,88,89,91,92,94],{},"Conceptually, the acoustic model essentially maps acoustic",[45,78],{},"\nobservations to the phone set.  The acoustic model can contain",[45,81],{},"\nentries for context independent elements from the phone set",[45,84],{},"\n(i.e., I don't care where the phone occurs) as well as",[45,87],{},"\ncontext dependent elements from the phone set (e.g., the",[45,90],{},"\n\"ah\" that occurs between \"b\" and \"r\" and is at the beginning",[45,93],{},"\nof a word).",[10,96,97,98,100,101,103,104,106,107,109],{},"The pronunciation lexicon maps words to sequences of elements",[45,99],{},"\nfrom the phone set.  For example, a phoneme-based system",[45,102],{},"\nmay give a pronunciation for \"eight\" as \"EY T\", whereas a",[45,105],{},"\none that uses word dependent phones may give the pronunciation",[45,108],{},"\nas \"EY_eight T_eight\".",[10,111,112,113,115,116,118,119,121,122,124,125,127],{},"The language model defines the allowable set of words that",[45,114],{},"\nwill be recognized and provides probability information about",[45,117],{},"\nsequences of words.  The language model can be complex",[45,120],{},"\n(e.g., a stochastic n-gram model) or simple (e.g., a simple",[45,123],{},"\nJSGF grammar).  For your application, I'd guess a JSGF",[45,126],{},"\ngrammar that you write would be the better way to go.",[10,129,130,131,133,134,136,137,139,140,142],{},"If you have knowledge of the domain ahead of time, you can",[45,132],{},"\ntailor the individual components to the task, providing",[45,135],{},"\nincreased recognition accuracy.  Creating an acoustic model",[45,138],{},"\nfor the specific task can yield great accuracy, but is",[45,141],{},"\ndifficult to do because it requires lots of training data.",[10,144,145,146,148,149,151,152,154,155,157,158,160,161,163,164,168,170],{},"We have several acoustic models for Sphinx-4.  The TIDIGITS",[45,147],{},"\nmodel (has word dependent phones just for digits) almost",[45,150],{},"\nmeets your requirements and gives GREAT continuous digits",[45,153],{},"\nrecognition, but doesn't have yes\u002Fno built in, so it probably",[45,156],{},"\nwill not suffice.  The WSJ model is a much more general",[45,159],{},"\nmodel, and will give you coverage for your task.  Since it",[45,162],{},"\nis much more general, however, it may perform ",[165,166,167],"em",{},"slightly",[45,169],{},"\npoorer in accuracy for your specific task.",[14,172,174],{"id":173},"see-also","See Also",[176,177,178,186],"ul",{},[179,180,181,185],"li",{},[35,182,183],{"href":183,"rel":184},"http:\u002F\u002Fcmusphinx.sourceforge.net\u002Fsphinx4\u002F",[39]," - Sphinx4 Main Page",[179,187,188,191],{},[35,189,37],{"href":37,"rel":190},[39]," - Overview of major Sphinx4 components",{"title":58,"searchDepth":193,"depth":193,"links":194},2,[195,196,197],{"id":16,"depth":193,"text":17},{"id":29,"depth":193,"text":30},{"id":173,"depth":193,"text":174},"2004-10-07T02:41-04:00",false,"md",null,{},true,"\u002Fsphinx-notes",{"title":5,"description":12},"Sphinx Notes",[],"mFRZIbiIxVZcdO15_tX99O8jz_iOCTR7SMGWufrjxnc",1791636469115]