ITU-T P 59-1993 ARTIFICIAL CONVERSATIONAL SPEECH《仿真会话语声》.pdf

1、ITU-T RECMN*P.59 93 W 4862593 0585bqb 3b5 W INTERNATIONAL TELECOMMUNICATION UNION ITU=T TELECOMMUNICATION STANDARDIZATION SECTOR OF ITU TELEPHONE TRANSMISSION QUALITY OBJECTIVE MEASURING APPARATUS P.59 (03193) ARTIFICIAL CONVERSATIONAL SPEECH ITU-T Recommendation P.59 (Previously “CCIlT Recommendati

2、on”) ITU-T RECflN*P.59 93 4862593 0585697 2T3 H FOREWORD The ITU Telecommunication Standardization Sector (ITU-T) is a permanent organ of the International Telecom- munication Union. The ITU-T is responsible for studying technical, operating and tariff questions and issuing Recommendations on them w

3、ith a view to standardizing telecommunications on a worldwide basis. The World Telecommunication Standardization Conference (WTSC), which meets every four years, established the topics for study by the ITU-T Study Groups which, in their turn, produce Recommendations on these topics. ITU-T Recommenda

4、tion P.59 was prepared by the IT-T Study Group XII (1988-1993) and was approved by the WTSC (Helsinki, March 1-12, 1993). NOTES 1 As a consequence of a reform process within the International Telecommunication Union (ITU), the CCITT ceased to exist as of 28 February 1993. In its place, the ITU Telec

5、ommunication Standardization Sector (ITU-T) was created as of 1 March 1993. Similarly, in this reform process, the CCIR and the IFRB have been replaced by the Radiocommunication Sector. In order not to delay publication of this Recommendation, no change has been made in the text to references contai

6、ning the acronyms “CCITT, CCIR or IFRB” or their associated entities such as Plenary Assembly, Secretariat, etc. Future editions of this Recommendation will contain the proper terminology related to the new ITU structure. 2 telecommunication administration and a recognized operating agency. In this

7、Recommendation, the expression “Administration” is used for conciseness to indicate both a O ITU 1994 All rights reserved. No part of this publication may be reproduced or utilized in any form or by any means, electronic or mechanical, including photocopying and microfilm, without permission in writ

8、ing from the ITU. ITU-T RECMNaP-57 73 = 4862573 0585b78 L38 CONTENTS Page 1 Introduction 1 2 Characteristics of human conversational speech 1 2.1 Characteristics measured without hangover time . 1 2.2 Characteristics measured with hangover time 2 3 Method of generating artificial conversational spee

9、ch. 3 References . 4 Recommendation P.59 (03/93) i ITU-T RECMN+P.59 93 = 48b259L 0585699 074 9 Recommendation P.59 ARTIFICIAL CONVERSATIONAL SPEECH (Helsinki, 1993) 1 Introduction The signal described here reproduces the on-off temporal characteristics of human conversational speech for characterizi

10、ng speech processing systems which have speech detectors, such as loudspeaker telephones, echo control devices, digital circuit multiplication equipment (DCME), packet systems, and asynchronous transfer mode (ATM) systems. This signal reflects parameters of human conversation such as the length of t

11、he talk-spurt, pause, double talk, and mutual silence. The following chapters describe these Characteristics and a method of generating artificial conversational speech. NOTES 1 The artificial voice described in RecommendationP.50 is a single-channel signal without pauses and is used for objective m

12、easurements of speech processing systems and devices in which the conversational environment is not relevant, such as speech codecs. The artificial conversational speech described in this Recommendation generates the artificial voice described in Recommendation P.50 during talk-spurts. 2 2 The durat

13、ions and rates of talk-spurt and pause vary according to the measurement conditions. The following specifies two values for each parameter in conversational speech. One is based on measurement of speech without hangover time, while the other is from that with hangover time. Characteristics of human

14、conversational speech 2.1 The characteristics described below were derived from Reference i. Characteristics measured without hangover time 1) Talk-spurt characteristics The probability density function (pdf) of talk-spurt duration is modelled by two weighted geometric pdfs: fi(k) = Cl(1 - U,) U!- +

15、 C2( l-U2 ) U5-l.k = 1,2,3, . where Cl = 0.60278 C, = 0.39817 U1 = 0.92446 U2 = 0.98916 Every increment of the variable k is equal to 5 ms. The cumulative distribution function of talk-spurt durations is shown in diagram a) of Figure 1. The average talk-spurt duration is 227 ms. 2) Pause characteris

16、tics The pdf of pause duration is also modelled by two weighted geometric pdfs: where Dl = 0.76693 Wi = 0.89700 D2 = 0.23307 W2 = 0.99791 The cumulative distribution function of pause duration is shown in diagram b) of Figure 1. Recommendation P.59 (03/93) 1 ITU-T RECMNxP.59 93 V8b2591i 0585700 blib

17、 w Parameter Duration (s) Talk-Spurt 1 .o04 Pause 1.587 Double talk 0.228 Mutual silence 0.508 3) Activity factor Rate (%) 38.53 61.47 6.59 22.48 The average pause duration of 596 ms, combined with the 227 ms average talk-spurt duration, yields a long-term speech activity factor of 27.6 per cent. Re

18、commendation P.56 is used, a higher activity factor is to be expected (see Table 1). NOTE - This value is measured by a meter without hangover. However, if a meter conforming to % % “ 1 5 10 50 100 500 1000 Number d frames “ 1 5 10 50 loo 500 1000 ,t,ff Tl202SC-91/L101 a) Talk-spwt dvatbn in 5 ms fr

19、ames b) Pause duratbn in 5 msframes FIGURE 1R.59 Cumulative distribution of talk-spurt and pause durations (without hangover time) 2.2 Characteristics measured with hangover time Table 1 lists the values of feature parameters in human conversational speech. These values were obtained by averaging th

20、e values reported in References 1-4. TABLE 1R.59 2 Recommendation P.59 (03/93) ITU-T RECMN*P-59 93 Y862591 0585701 552 W The cumulative distribution function of talk-spurt duration is approximated by an exponential function and that of pause durations is approximated by a constant-plus-exponential.

21、That is, for talk-spurt: Pts(t) = 1 - exp(-Ats . t) Ats = ifits, Tts: average talk-spurt duration, and for pause, O 1 - exp-Aps(t - 0.2) for0 5 t I 0.2 fort 0.2 Aps = l/(ips - 0.2) Tps: average pause duration. Both characteristics are shown in Figure 2. % YO “1 5 10 50 100 500 loo0 Number of frames

22、a) Talk-sputl dvation in 5 ms frames b) Pause duration in 5 ms frames FIGURE 2P.59 Cumulative distribution of talk-spurt and pause durations (with hangover time) 3 Method of generating artificial conversational speech Talk-spurts and pauses are generated according to the state transition model shown

23、 in Figure 3, in which Pl, P2, and P3 denote transition probabilities expressed in per cent. The artificial voice described in Recommendation P.50 is generated during a talk-spurt. Tst (single talk duration), Tdt (double talk duration), and Tms (mutual silence duration) vary according to the followi

24、ng equations. The times in these equations are expressed in seconds. Tst = -0.854In(l - xi) Tdt = -0.2261n(l - x2) Tmc = -0.4561n(l - x3) O e XI, x2. x3 e 1: Random variables with uniform distribution. Recommendation P.59 (03/93) 3 ITU-T RECMN*P.57 73 = 4862573 0585702 q79 = If the pause duration is

25、 less than 200 ms, the model chooses either the single talk or mutual silence state with probabilities of 50% until the pause duration exceeds 200 ms. The values of Pl, P2, and P3 are 40, 50, and 50, respectively. The total duration of artificial conversational speech must be at least 10 minutes to

26、comply with the Characteristics specified in 2.2. Talkspurt t A t B I Talkspurt Pause Single talk A: Silence B: Talk T1203840-91/603 _ FIGURE 3P.59 State transition model for conversation References l LEE (H.H.), UN (C.K.): A study of on-off characteristics of conversational speech, ZEEE Trans. on C

27、omm., Volume COM-34, No. 6, pp. 630-637, June 1986. BRADY (P.T.): A statistical analysis of on-off patterns in 16 conversations, BSTJ, pp. 73-91, January 1968. CCIIT Contribution COM XII-20, On-off characteristics of conversational speech (CSELT), Study Period 1989-1992. CCIIT Contribution Delayed D.42 (WP XIYl), Collecting procedure for on-off characteristics of conversational speech in telecommunication (NIT), Study Period 1989- 1992. CCITT Contribution Delayed COM-64 (WP XIUl), Generation of artifcial voice with pauses (NTT), Study Period 1989-1992. 2 3 4 5 4 Recommendation P.59 (03/93)