|
|
Detekce změny jazyka při hovoru
Povolný, Filip ; Glembek, Ondřej (oponent) ; Matějka, Pavel (vedoucí práce)
Tato práce se zabývá problematikou detekce změny jazyka při hovoru. V první části jsou popsány v současnosti používané metody diarizace jazyků. K implementaci byla vybrána metoda založená na akustickém přístupu identifikace jazyka s využitím směsi Gaussovských rozložení, i-vektoru a lineární diskriminační analýzy. Pro experimenty byla vytvořena mandarínsko-anglická databáze se střídáním jazyků. Na této databázi zvolený systém dosahuje úspěšnosti 89,3 % správně klasifikovaných segmentů.
|
|
|
PHONOTACTIC AND ACOUSTIC LANGUAGE RECOGNITION
Matějka, Pavel ; Sigmund, Milan (vedoucí práce)
This thesis deals with phonotactic and acoustic techniques for automatic language recognition (LRE). The first part of the thesis deals with the phonotactic language recognition based on co-occurrences of phone sequences in speech. A thorough study of phone recognition as tokenization technique for LRE is done, with focus on the amounts of training data for phone recognizer and on the combination of phone recognizers trained on several language (Parallel Phone Recognition followed by Language Model - PPRLM). The thesis also deals with novel technique of anti-models in PPRLM and investigates into using phone lattices instead of strings. The work on phonotactic approach is concluded by a comparison of classical n-gram modeling techniques and binary decision trees. The acoustic LRE was addressed too, with the main focus on discriminative techniques for training target language acoustic models and on initial (but successful) experiments with removing channel dependencies. We have also investigated into the fusion of phonotactic and acoustic approaches. All experiments were performed on standard data from NIST 2003, 2005 and 2007 evaluations so that the results are directly comparable to other laboratories in the LRE community. With the above mentioned techniques, the fused systems defined the state-of-the-art in the LRE field and reached excellent results in NIST evaluations.
|
|
|
Detekce změny jazyka při hovoru
Povolný, Filip ; Glembek, Ondřej (oponent) ; Matějka, Pavel (vedoucí práce)
Tato práce se zabývá problematikou detekce změny jazyka při hovoru. V první části jsou popsány v současnosti používané metody diarizace jazyků. K implementaci byla vybrána metoda založená na akustickém přístupu identifikace jazyka s využitím směsi Gaussovských rozložení, i-vektoru a lineární diskriminační analýzy. Pro experimenty byla vytvořena mandarínsko-anglická databáze se střídáním jazyků. Na této databázi zvolený systém dosahuje úspěšnosti 89,3 % správně klasifikovaných segmentů.
|
|
|
PHONOTACTIC AND ACOUSTIC LANGUAGE RECOGNITION
Matějka, Pavel ; Sigmund, Milan (vedoucí práce)
This thesis deals with phonotactic and acoustic techniques for automatic language recognition (LRE). The first part of the thesis deals with the phonotactic language recognition based on co-occurrences of phone sequences in speech. A thorough study of phone recognition as tokenization technique for LRE is done, with focus on the amounts of training data for phone recognizer and on the combination of phone recognizers trained on several language (Parallel Phone Recognition followed by Language Model - PPRLM). The thesis also deals with novel technique of anti-models in PPRLM and investigates into using phone lattices instead of strings. The work on phonotactic approach is concluded by a comparison of classical n-gram modeling techniques and binary decision trees. The acoustic LRE was addressed too, with the main focus on discriminative techniques for training target language acoustic models and on initial (but successful) experiments with removing channel dependencies. We have also investigated into the fusion of phonotactic and acoustic approaches. All experiments were performed on standard data from NIST 2003, 2005 and 2007 evaluations so that the results are directly comparable to other laboratories in the LRE community. With the above mentioned techniques, the fused systems defined the state-of-the-art in the LRE field and reached excellent results in NIST evaluations.
|
host ::
RSS.