TOWARDS HUMAN TRANSLATIONS GUIDED LANGUAGE DISCOVERY FOR ASR SYSTEMS
Sebastian Stüker, Alex Waibel · 2008
ABSTRACTNatural language processing systems, e.g for AutomaticSpeech Recognition (ASR) or Machine Translation (MT),havebeenstudiedonlyforafractionoftheapprox.7000lan-guagesthatexistintoday’sworld,themajorityofwhichhaveonlycomparativelyfewspeakersandfewresources.Thetra-ditionalapproachofcollectingandannotatingthenecessarytrainingdataisduetoeconomicconstraintsnotfeasibleformostofthem. AtthesametimeitisofvitalinteresttohaveNLPsystemsaddresspracticallyalllanguagesintheworld.New,efficientwaysofgatheringtheneededtrainingmaterialhavetobefound.InthispaperweproposeanewtechniqueofcollectingsuchdatabyexploitingtheknowledgegainedfromHumansimultaneoustranslationsthathappenfrequentlyintherealworld. Toshowthefeasibilityofourapproachwepresentfirstexperimentstowardsconstructingapronuncia-tiondictionaryfromthedatagained.Index Terms — Automatic Speech Recognition, Lan-guage Discovery, Machine Translation, Under-ResourcedLanguages1. INTRODUCTION1.1. The Traditional Way to Acquire Training DataTraining large vocabulary continuous speech recognition(LVCSR) systems requires a number resources in the tar-getedlanguage. Fortrainingtheacousticmodelofarecog-nitionsystemlargeamountsoftranscribedaudiorecordingsofspeechareneeded.Thetrainingofthelanguagemodelre-quireslargeamountsofwrittentextinthetargetedlanguage.Whenusingphonemebasedacousticmodels,apronunciationdictionaryisneededthatmapsthewrittenrepresentationofawordtothesequenceofitsphonemeswhenbeingspoken.Approximately7,000languagesexisttoday,thecurrenteditionofEthnologue[1]lists7,299.Sofar,automaticspeechrecognition (ASR) systems and machine translation (MT)