Presentation is loading. Please wait.

Presentation is loading. Please wait.

Detecting Accent Sandhi in Japanese Using a Superpositional F0 Model Atsuhiro Sakurai Hiromichi Kawanami Keikichi Hirose Depart. of Communication and Information.

Similar presentations


Presentation on theme: "Detecting Accent Sandhi in Japanese Using a Superpositional F0 Model Atsuhiro Sakurai Hiromichi Kawanami Keikichi Hirose Depart. of Communication and Information."— Presentation transcript:

1 Detecting Accent Sandhi in Japanese Using a Superpositional F0 Model Atsuhiro Sakurai Hiromichi Kawanami Keikichi Hirose Depart. of Communication and Information Engineering The Univ. of Tokyo, JAPAN

2 To propose an algorithm that automatically detects the accent sandhi pattern of a Japanese compound noun, based on a superpositional F0 model Objective Automatic labeling using the F0 model can be useful for designing a prosodic database Accent sandhi in compound nouns (especially with 3 or more nouns) is a complex phenomenon Background

3 Outline Accent Sandhi and Accent Sandhi Pattern Detecting the accent sandhi type of two- word compound nouns Detecting the accent sandhi pattern of compound nouns containing more than 2 words

4 Accent Sandhi When several nouns merge to form a compound noun, the original accent nuclei of the component nouns change their positions or disappear. We propose a method to automatically analyze the accent sandhi phenomenon and test it in two cases: compound containing 2 nouns, and those containing more than 2 nouns.

5 Detecting Accent Sandhi for 2- Word Compound Nouns According to NHK Pronunciation and Accent Dictionary, the shape of 2-noun compound nouns is determined by the second component. The 2nd component noun can be classified into 4 types.

6 Accent Sandhi Patterns (According to the 2nd component noun:) Type A : nucleus at first mora of second noun ( Example: “asobi” + “aite” = “asobia’ite” ) Type B : nucleus at last mora of first noun ( Example: “seifu” + “aN” = “seifu’aN” ) Type B* : nucleus at penultimate mora of first noun ( Example: “geNzei” + “aN” = “genze’iaN” ) Type F: flat ( Example: “akita” + “keN” = “akitakeN” )

7 Phoneme Labels and timing Type A Type B Type B* Type F F0 Contour Hypothesizer Model A Model B Model B* Model F Partial Abs Error A Error B Error B* Error F System Outline Error = MSE between extracted and calculated F0 contours

8 F0 Contour Model

9 Approximate Model for Compound Nouns ( Initial Values ) 1.00.08 t (s) Command A p1 A p2 A a1 t 01 t 02 t1t1 t2t2 By using 2 phrase commands, all possible prosodic structures can be simulated After phrase boundaries with reset: (Ap1=0,Ap2>0) After other phrase boundaries: (Ap1>0,Ap2>0) After non-phrasal boundaries: (Ap1>0,Ap2=0)

10 Initial Values of Timing Parameters h a n a sh i k o t o b a -70 ms t1t1 t2t2 (for type A) -70 ms t2t2 (for type B) -70 ms t2 (for type B*)

11 Parameter Optimization Using Partial AbS Rough adjustment Fine tuning Initial values of timing parameters Calculation of error with respect to measured F0 contour ( Only phrase command magnitudes and accent command amplitude) ( All parameters )

12 Rough Parameter Adjustment for (A p1 =0.0; A p1 <=0.8; A p1 +=0.05) for(A p2 =0.0; A p2 <=0.8; A p2 +=0.05) { Calculate(A a ); if(distance<min) min=distance; } (A p1 *,A p2 *,A a *)

13 Parameter Fine Tuning A p1* A p2* A a* t 01 t 02 t1t1 t2t2 Order: (±20%) (±20 ms) 1) Phrase command magnitudes (A p1, A p2 ) 2) Phrase command times (t 01, t 02 ) 3) Accent command amplitude (A a ) 4) Accent command times (t 1, t 2 )

14 Evaluation Tests Speech material : ATR Continuous Speech Database ( MAU and MHT) Phoneme labeling by HTK speech recognizer in forced alignment mode

15 (a) Speech waveform (b) Phoneme labels (c) F0 contour (d) Model for type A (e) Model for type F Example of automatic accent sandhi type detection

16 Accent Sandhi Pattern of Long Compound Nouns Accent sandhi pattern = how component words concatenate to form new accentual phrases. For longer compound nouns, accent sandhi becomes harder to predict We extended the present method to detect accent sandhi patterns of long compound nouns (containing more than 3 nouns).

17 Accent Sandhi Pattern of Long Compound Nouns H1:So’oru goriNkoohose’Nshu H1’: SoorugoriN koohose’Nshu H2: ChuugokujiNuNte’Nshu H2’: ChuugokujiN uNte’Nshu S1 S2 Two sentences (S1 and S2) spoken each one by two individuals (I1 and I1’ for S1, I2 and I2’ for S2) using each one a different accent sandhi pattern (I1 uses H1, I1’ uses H1’, I2 uses H2, and I2’ uses H2’).

18 Accent Sandhi Pattern of Long Compound Nouns S o o r u g o r i N k o o h o s e N sh u C h u u g o k u j i N u N t e N sh u H1: H1’: H2: H2’:

19 Accent Sandhi Pattern of Long Compound Nouns 0.00 0.50 1.00 1.50 2.00 2.50 3.00 3.50 Correct Incorrect AbS Error (x 10 -2 ) I1I1’I2 I2’ H1 H1’ H1 H1’ H2 H2’ H2 H2’

20 Comments Present method works when the position of the accent nucleus on the F0 contour is visually clear. Difficult at long unvoiced segments ( “himitsu-kikai”, etc. ) Automatic labeling was one of the causes of errors.


Download ppt "Detecting Accent Sandhi in Japanese Using a Superpositional F0 Model Atsuhiro Sakurai Hiromichi Kawanami Keikichi Hirose Depart. of Communication and Information."

Similar presentations


Ads by Google