Razpoznavanje slovenskega govora z metodami globokih nevronskih mrež

Ulčar, Matej (author), Robnik Šikonja, Marko (mentor), Dobrišek, Simon (co-mentor)

Ročno zapisovanje govora je počasen proces, ki ga čedalje bolj nadomešča avtomatsko razpoznavanje govora. Slednje se lahko uporablja tudi za glasovno upravljanje programov in naprav. V magistrski nalogi smo kot osnovo za razpoznavanje govorjene slovenščine uporabili uveljavljene metode GMM-HMM za akustični model in n-gramov za jezikovni model. Modela smo nadgradili z uporabo globokih nevronskih mrež, ki so se izkazale za zelo uspešne. Preizkusili smo različne arhitekture časovno zakasnjenih nevronskih mrež in nevronskih mrež z dolgim kratkoročnim spominom na akustičnem in jezikovnem modelu razpoznavalnika govora. Razpoznavalnik smo učili na širokem besednjaku, ki vsebuje približno milijon različnih besed. Najboljše rezultate dosegajo časovno zakasnjene nevronske mreže, kjer smo dosegli 72,84% pravilno prepoznanih besed pri tekočem govoru.

• Language: Slovenian
• Keywords: strojno učenje, globoke nevronske mreže, razpoznavanje govora
• Work type: Master's thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2018
• PID: 20.500.12556/RUL-104850
• Publication date in RUL: 12.10.2018

Secondary language

• Language: English
• Title: Computer Speech Recognition in Slovene Language

Abstract

Manual transcription of speech is slow and is being replaced by automatic speech recognition systems. These systems are also used for voice control of various programs and devices. In this thesis, we used as a baseline for Slovene speech recognition GMM-HMM methods for acoustic model and n-grams for language model. We improved both models with deep neural networks, which have proven to be very successful. We tested several architectures of time-delayed neural networks and neural networks with long short-term memory for both acoustic and language model. We used a large lexicon, containing about a million words. Time-delayed neural networks achieved the best results on continuous speech, with 72,84% of correctly identified words.

• Keywords: machine learning, deep neural networks, speech recognition