Speaking Styles Advanced Signal Processing SE WS 03/04 Robin Hofe Contents Definition of Speaking Styles Applications Methodology Featured Investigations: Hyperactive Articulation Fast and Slow Speech Spontaneous and Read Speech Speaking formants Fun Speaking Style Prosody Variation with Text Type
The LAIPTTS System References Definition of Speaking Styles 1/2 In literature, there is so far no standard framework for classifying speaking styles. Here, I choose the following definitions to specify my work: Speaking style is a set of properties by which you can link speech to a certain communicative situation. This situation is multi-dimensional: - Content (news, poem, dialogue, etc.) - Speaker
(habits, personality, etc.) - Situation (distance, noise, room size, etc.) - Listener (relation to speaker, number, etc.) Definition of Speaking Styles 2/2 In the following chapters, we often have a pair of characteristics which will be compared (e.g. fast - slow). This is obviously not the best solution but it can give you clear results. As this is a very young research field and most featured studies are pilot projects, I can only give quantitative
results. I have excluded emotional speaking styles. Applications Tutoring systems, dialogue systems Medical applications, nursing robots Animated characters Talking newspapers for the blind Applications in hands-busy-eyes-busy situations Various applications in voice recognition Methodology How do we collect relevant data ? Production of spontaneous speech in the laboratory Do professional speakers produce representative results of common speech ? What do I examine in the signal (f0, formants, durations,
etc.) ? How will the data be processed ? Use of the procedures already introduced in previous talks. Which procedures are optimal ? Featured Investigations In the following chapters the structure will be: Authors Aim of the investigation Database Methods applied Results and discussion It is important to point out that these investigations were carried out in certain languages and apply without further studies to those only!
Hyperactive Articulation 1/7 Stefanie Kster: Acoustic-phonetic Characteristics of Hyperarticulated Speech for Different Speaking Styles Aim: To describe the difference between normal and hyperarticulated speech. Parameters: duration, fundamental frequency, formants, formant bandwidth. Database: German speech, normal and hyperarticulated. Isolated words, sentences and spontaneous speech. 3 female and 3 male speakers. Methodology: For normal speech, the speaker had to address a person in front of him/her. For hyperarticulated speech, this person had headphones, signalling a communication obstacle. Hyperactive Articulation 2/7 Methodology (2): The spontaneous speech was created by simulating a train reservation system.
Furthermore, an auditory test was performed to find perceptional differences between the two speaking styles: for each stimulus, the listeners had to decide how it applies to a pair of antonyms on a scale from 0 (e.g. slow) to 6 (e.g. fast). Hyperactive Articulation 3/7 Results: Significant differences can be observed between the two speaking styles: Duration increases for hyperarticulated speech in all 3 text types. Figure 1: Duration increases for all text styles after eliminating all pauses. Table 1: Average duration change [%] for the phoneme classes.
> Vowels seem to be important. Hyperactive Articulation 4/7 Figure 2: Average change in the duration of vowels depending on their position in the sentence. > Emphasis on the beginning and ending of a sentence seems to be important. Hyperactive Articulation 5/7 Fundamental Frequency: Increase visible in all three categories: The biggest difference is visible for single words: 25%. Spontaneous speech and sentences have 21.5% and
21.2% f0 increase. > Great rise in the variation of f0 within spontaneous speech. Hyperactive Articulation 6/7 Formants and formant bandwidth: Formant bandwidths are generally lower, formant frequencies decrease: Table 2: Differences [%] for all phoneme classes. > Sentences seem less affected. > The tendency of decreasing formant frequencies was very strong with fricatives. Figure 4: F1-F2 plane for long vowels, both hyperarticulated and normal speech.
> Different results have been found for English language! Hyperactive Articulation 7/7 Perceptional Differences: > Hyperarticulated speech is perceived as fast or slow as normal speech. > It was judged as uncomfortable, aggressive and powerful. This might well be a problem when it comes to synthesis. Fast and Slow Speech 1/6 Alex Monaghan: An Auditory Analysis of the Prosody of Fast and Slow Speech Styles in English, Dutch and German
Aim: Determining differences in accent location, boundary location and boundary strength at different speech speeds and whether there is consistency in those across languages. Database: The multilingual speech database, recorded in April 1999 as part of the COST 258 programme. Texts of the same general style (informative) but not of the same content, read in two speeds. COST: European Cooperation in the field of Scientific and Technical Research Fast and Slow Speech 2/6 Database (2): Languages and speakers were: English / Male Dutch / Female Austrian German / Male (GermanA)
Swiss German / Male (GermanS) Leipzig German / Female (GermanL) Bonn German / Female (GermanB) GermanS and GermanL had the same text. Fast and Slow Speech 3/6 Methodology: Manual transcription of three aspects of the prosody of the recorded speech by adding diacritics to the written text: accent location, boundary location, boundary strength. Boundary strength was transcribed using three categories: major pause (Utt) minor pause (IP) boundary tone, no pause (T) Fast and Slow Speech 4/6
Results: Table 1: Length in words, and duration of the six text samples Table 1 gives an overview over the different text samples: they vary greatly in length and speech speed difference. Accents: Words Fast Slow Fast/Slow
English 35 11,5s 17,0s 0,68 Dutch 75 24,0s
42,0s 0,57 GermanA 148 54,5s 73,0s 0,75 GermanB 78
28,0s 51,0s 0,55 GermanL 63 27,0s 49,0s 0,55
GermanS 63 25,5s 38,5s 0,66 There are never more accents in the fast version. The accents in the fast version occur only at locations, where there are accents in the slow version. > Fast Speech can be characterized by accent deletion.
Table2: Accents transcribed and Overlap between fast and slow variety Fast Slow Overlap English 21 21 21 (100%)
33 36 33 (92%) Fast and Slow Speech 5/6 Boundaries: Table 3: Total number of boundaries, all categories Total Boundaries As with accents, an increase of the number of boundaries with slow speech rate is visible. The lowest raise (30%, GermanA)
corresponds to the lowest decrease in speaking time. Table 4 shows a trend towards weakening of pauses with increasing speech speed. > Fast speech seems to be characterised by a decrease or deletion of boundary strengths. Fast Slow English
8 11 Dutch 10 21 GermanA 22 29 GermanB
12 36 GermanL 7 23 GermanS 8 15
Table 4: Changes in Boundary strength from slow to fast -2 -1 0 +1 +2 English 2 7
2 0 0 Dutch 5 14 2 0
0 GermanA 2 15 12 0 0 GermanB 11
17 8 0 0 GermanL 8 15 0
0 0 GermanS 3 10 2 0 0 Fast and Slow Speech 6/6
Important for TTS: In slow speech, Utt and IP boundaries correspond to punctuations and commas but T boundaries are not predictable this way. It appears that for a given text, different speakers agree on which accents to delete or boundaries to demote on a faster speech rate. Conclusion: For a given text and speech rate, accent location, Utt and IP boundary locations and strengths are predictable from the text. Faster speech rates can be constructed from slower ones. T boundaries remain unpredictable. Another investigation by Brigitte Zellner comes to the same conclusions for French. Spontaneous and Read Speech 1/10
Marc Swerts, Eva Strangert, Mattias Heldner: F0 Declination in Read-aloud and Spontaneous Speech Aim: Investigation of differences in the f0 contour between spontaneous and read Swedish speech. Database: 2 Swedish Speech samples spoken by a male person: A read-aloud news telegram (233 words) and a spontaneous retelling of the text (252 words). Methodology: Estimation of the slope of declination in the f0 contour by fitting an all-points regression line to the f0 points. The boundaries between phrases and utterances were determined by the mean scores of 9 transcribers. Spontaneous and Read Speech 2/10 Results: > The mean f0 was higher in readaloud (136 Hz) than spontaneous speech (107 Hz).
> Slopes were significantly steeper in the read-aloud version. > Duration was shorter for readaloud speech. Figure 1: Left: read-aloud, Right: spontaneous Spontaneous and Read Speech 3/10 The data from Table 1 suggests a time dependency of the f0 slope. Correlation coefficients between the length of phrases/utterances and the slope were calculated: > Significant correlations are given. > Speaker seems to adapt his speaking style due to the overview he has over the phrases in
the near future. Spontaneous and Read Speech 4/10 Danielle Duez: Reduction and Assimilatory Processes in Conversational French Speech Implications for Speech Synthesis Aim: To give an overview over effects that happen in conversational French compared to hyper articulated French. He uses examples of widely known conversational French and analyzes them linguistically. In the following, I will shorten or even omit his explanations and stick to the conclusions he draws from the examples. Spontaneous and Read Speech 5/10
Reduction and contextual assimilation: Voiced Stops: (1) There is a partial or complete nasalization of /b/ and /d/ proceeding or succeeding a nasal vowel. Example: /pd/ /pn/ Spontaneous and Read Speech 6/10 (2) Weakening of /b/ into // and /w/ and of /d/ into /z/ or /l/ or its complete deletion. These changes are assumed to result from a reduction in the magnitude of the closing gesture. Note that the place of articulation stays the same. Consonant Sequences: [C1#C2] or [C1C2], where # marks a syllable boundary. Often, C1s were changed into another consonant
or omitted. Voiced/unvoiced fricatives and plosives were devoiced/voiced due to the anticipatory effect of an unvoiced/voiced C2. Spontaneous and Read Speech 7/10 Examples: /ilmtataRive/ /imtataRive/ Omission of the /l/. /yntasptasd/ /yntasptaz/ Fusion of /sd/ to /z/. Factors limiting reduction and assimilation effects: Segment properties: Dentals show a higher resistance against reduction and assimilatory effects than labials. Sonorants were the consonants that were omitted most frequently.
Spontaneous and Read Speech 8/10 Voiced occlusives are less resistant than unvoiced ones, having less articulatory force. The more resistant segments could in turn be the ones with stronger influence on their neighbours. Syllable structure: Highest identification score for fricative-sonorant sequences, lowest for occlusive-occlusive. Higher identification rate for homosyllabic consonant sequences than for heterosyllabic. Spontaneous and Read Speech 9/10 Word Class: Elided article or preposition is often omitted or changed to /n/. The of often changes or removes the first consonant of a following verb.
Tendencies in reduction and assimilation: Nasalization: Nasalization in French tends to spread from one segment to another. Spontaneous and Read Speech 10/10 Resistance of prominent final-phrase syllables: There seem to be important landmarks, which attract the listeners special attention and are crucial for comprehensibility. Speaking Formants 1/3 Gunilla C. Thunberg: Balancing Spectra Between Different Speaking Styles Aim: Exploring the use of spectral balance in different speaking styles: Is there a speakers formant? Database: 2 female professional actors performed the
same text by first using their stage voice and then a conversational tone for close range. Methodology: Recordings under rehearsal-like conditions on a stage. Spectrum analysis was applied to the recordings. Speaking Formants 2/3 Results: SubjectB (left) has higher f0 values altogether. Speaking Formants 3/3 Discussion: SubjectB has apparently no change in the spectral contour. While SubjectG could be compared to an alto singer, SubjectB is more like a soprano. Due to their very high f0 and thereby fewer
overtones, sopranos do generally not use a singers formant. SubjectG shows 3 extra peaks: at 2-3 kHz (where the singers formant would be) but also at 4-5 kHz and 5-6 kHz. Further investigation will be necessary. Fun Speaking Style 1/6 Kjell Gustafson, David House: Prosodic Parameters of a Fun Speaking Style Aim: To examine how children react to prosodic variations differing from the standard TTS for adult audiences. Database: 3 sentences, spoken by the Infovox 230 formant-based Swedish male voice and the Infovox 330 concatenated diphone Swedish female voice. Four different versions of each sentence were
produced: (1) standard, (2) double duration in focused words, (3) double of maximum f0 values in focused words, (4) combination of (2) and (3). Fun Speaking Style 2/6 Methodology: On a computer, an animated character was shown (an astronaut). 8 text fields next to it were linked to sound files. The subjects could activate those by clicking on them. They were then asked, which example they found the most natural one and which the most fun. Subjects were 8 children and 4 adults as a control group. Fun Speaking Style 3/6 Results:
> Large manipulation of both f0 and duration were preferred for a fun voice. > For naturalness, children preferred higher f0 peaks. > The f0 manipulations were found to be more funny by both adults and children. > Children distinguish strongly between naturalness and fun. Fun Speaking Style 4/6 This figure shows the result of a ranking task for the three sentences. > Extreme prosody is clearly preferred in both naturalness
and fun. The children preferred the formant synthesis over the diphone synthesis, the reason could be that it was seen as more adequate for an animated character. For a larger character (lion, elephant), a low f0 might be preferred. Fun Speaking Style 5/6 Sound examples: The sentence presented is: Idag ska jag flyga till en ANNAN planet. Today Im going to fly to a DIFFERENT planet. Formant Synthesizer, default Diphone Synthesizer, default
Formant Synthesizer, f0 Diphone Synthesizer, f0 Formant Synthesizer, duration Diphone Synthesizer, duration Formant Synthesizer, f0+duration Diphone Synthesizer, f0+duration Fun Speaking Style 6/6 Tasks for future studies: Investigate ideal values for the modifications and optimum shapes of the f0 contour. Should the modifications be restricted to focused words?
Is the effect the same with longer texts? Prosody Variation with Text Type 1/7 The Variation of Prosody with Text Type, Justin Fackrell, Halewijn Vereecken, Jean-Pierre Martens, Bert Van Coile Aim: Examination of text type dependent variations of prosody on a document lever rather than on a sentence or word level. Database: Texts from 10 text types in three languages (Dutch [Belgium], English [US] and French [France]) were collected from the internet and read by female native speakers. (3 hours of material) Methodology: Word prominence (PRM) and inter-word prosodic boundary strength (PBS) were transcribed using a 0..5 scale.
Prosody Variation with Text Type 2/7 Methodology (2): A pitch extractor determined a pitch track for each recording. The grapheme-to-phoneme component of the RealSpeakTM TTS engine was used to get a prototypical phonetic transcription of the texts. Word length in syllables and average phrase length in syllables were measured. Prosody Variation with Text Type 3/7 Results: Figure 1: Showing the word length variations from one text type to another. > Childrens stories (but also adults
stories) have the shortest, CVs the longest words. > Over all, English has shorter words. Prosody Variation with Text Type 4/7 Figures 2 and 3 show the speaking rate in syllables/minute and the difference from the average speed of the language per text type. There are great differences between the text types, but also between the speakers reading the same text type. Prosody Variation with Text Type 5/7 Figure 4 shows a scatterplot between median phrase and pause duration. Some text types show consistency over languages.
Prosody Variation with Text Type 6/7 Figures 5 and 6 show the variation of average pitch depending on text type and language. There are, again, differences between languages and speakers. Prosody Variation with Text Type 7/7 Discussion: Prosody variation seems to be at least language dependant, if not speaker dependant. (Probably both) The strongly diverse text types seemed to be CVs and childrens stories. The LAIPTTS System 1/5 A Nonlinear Rhythmic Component in Various Styles of
Speech, Brigitte Zellner Keller, Eric Keller Goal: Developing a dynamic model of the temporal organisation of speech. This requires a complex nonlinear dynamic model. Flaws of current approaches: Assumption of a linear relation between variables influencing speech timing. The variability is underestimated. Human speech gestures are never repeated exactly the same way. Statistical approaches often lead to a false impression of causal relation. The LAIPTTS System 2/5 The BioPsychoSocial Speech Timing Model: The levels of constraints that govern speech activity in the time domain: (1) Bio-psychological: respiration, neurology,
psycho-rhythmic tendencies. (2) Social: linguistic and socio-linguistic constraints. (3) Pragmatic: situation, feelings, cognitive tasks. (1) is the base level, on which (2) and (3) superimpose their constraints. The LAIPTTS System 3/5 More effort should be put into the investigation of serial temporal dependencies. Autocorrelation methods should be applied to find dependencies of an Xk+1|X1..Xk type. Applying such methods, the LAIP found regular anticorrelations for French and, to a lesser extend, English. for
The anticorrelation component manifested itself reliably within 500ms or 1 or 2 syllables. This could have a neurophysiological and/or articulatory cause. It may even be a general human rhythmic tendency. The LAIPTTS System 4/5 This anticorrelational effect was implemented as a parameter in the LAIPTTS Speech Mill speech synthesis system and added a swingy effect with high values or a more controlled effect with weak ones. The LAIPTTS SpeechMill uses a grapheme-to-phoneme translator, the LAIP-designed prosody control system for timing and f0 and Mbrola diphone motor. The LAIPTTS System 5/5
Sound examples: Welcome Unmarked reading language Normal speed Slow speed Fast speed Explanation of the LAIP project Beat Siebenhaar wohnt mit seiner Familie in der Kleinstadt Aarau, die zwischen Zrich und Bern liegt. Fr die Fahrt zur Arbeit nimmt er die Bahn, die ihn mit direkten Verbindungen in zweieinhalb Stunden nach Lausanne bringt. Wenn er nicht das Familienleben mit seinen 3 Tchtern pflegt oder am Computer sitzt, fhrt er manchmal seine 30-jhrige goldene DS
spazieren. Resources 1/2 > Improvements in Speech Synthesis Edited by E. Keller et al. 2002 http://www2.unil.ch/imm/cost258volume/cost258volume.htm > Acoustic-phonetic Characteristics of Hyperarticulated Speech for Different Speaking Styles, Stefanie Kster, Ruhr University Bochum > Temporal Structures for Fast and Slow Speech Rate, Brigitte Zellner, LAIP, University of Lausanne > F0 Declination in Read-aloud and Spontaneous Speech, Marc Swerts, Eva Strangert, Mattias Heldner > Balancing Spectra between Different Speaking Styles, Gunilla C. Thunberg, Stockholm University > The Variation of Prosody with Text Type, Justin Fackrell, Halewijn Vereecken, Jean-Pierre Martens, Bert Van Coile
The group 1A elements, called the alkali metals, are all reactive metals. The group 2A elements are called the alkalineearth metals: fairly reactive, but not quite as reactive as the alkali metals. Prediction of New Elements. ... The Periodic Law
Core concepts and the info explosion. In organizing and planning our courses, a focus on the core concepts of physiology may provide at least a partial solution to the problem of students learning all they need to learn in the...
IT Steering Security Subcommittee Status Update December 1, 2011 Electronic device security policy Virtual Private Network (VPN) software Securely connect to the U from anywhere Software Depot download (free) Windows 7, XP, Mac, Linux versions available Electronic device security policy...
This is different from the existing split liver variance in that either segment can be used for the first candidate. Under the existing variance, participating hospitals can only use the second segment at their own center if the first candidate...
Indirect Acting Cholinomimetics Pharmacodynamics: The affinity of the drug to acetycholinesterase determines the duration of action edrophonium and related quaternary alcohols interact weakly (electrostatic and hydrogen bonds) --> 2-10 min interaction carbamate esters (e.g., neostigmine) form ...
Safety Decision Making Methodology (1 of 6) Intake through closure. INFORMATION DOMAINS: Making good decisions about safety begins with gathering sufficient information using the following six standardized information domains throughout the life of a case.
Lay the groundwork for both the state-level Policy Academy work and the cross-state Action Network efforts. ... Referral and Service Coordination. Discipline Policy Consultation. ... Understanding and Partnering with the Juvenile Justice System and Local Law Enforcement.
Chebucto Plus 56 K dialup service: $100 per year Sustaining membership includes as much access as you need Mailbox accessible from anywhere using webmail Membership Reward Plan saves you money for each new user you refer Family plans available for...
Ready to download the document? Go ahead and hit continue!