Intonation of telephone conversations in a Customer Care Service RATP-DECODA is a Customer Care Service corpus part of the ORFEO project (2020). It includes 1988 recordings of client requests made between 2009 and 2011 to the RATP call center (Régie Autonome des Transports Parisiens). The total duration of these calls is about 98 hours, with an average of 177 seconds per call. Recordings were made in a mp3 format, converted in the Orfeo corpus into stereo 16 bit 8000 Hz sampling rate, and for the present research in mono 16 bit 22050 Hz. The files are originally annotated in word and phone segments, in trs (Transcriber) and also into TextGrid (Praat) format. The compressed recording format used (mp3) and the telephone line characteristics limit the sound bandwidth to about 120-3900 Hz, which affects the reliability of most pitch trackers and of the automatic segmentation program used for word and phone annotations, especially for male voices. Nevertheless, using dedicated functions of the WinPitch (2020) speech analysis software, errors of segmentation were easily corrected manually, and pitch detection errors have been monitored by the simultaneous display of a narrow band spectrogram and graphically corrected (Martin, 2018a). Conversations usually involve 2 participants: a client, the customer agent and eventually a service voice. Call types as noted in Brechet et al. (2012) were: 
Info Traffic 22.5%; Route planning 17.2%; Lost and Found 15.9%; Registration card 11.4; Timetable 4.6%; Ticket 4.5%; Specialized calls 4.5%; empty 3.6%; New registration 3.4%; Price info 3.0%.
While some semantic and syntactic observations have already be made in Brechet et al. (2012), no analysis of intonation has so far been conducted, showing in these specific discourse conditions the role of prosody in structuring controlled exchanges between two participants. As most of the recordings have a duration below 3 minutes, the effect of client characteristics on the agents, for instance perceived age and fluency, can be analyzed more easily in this restricted time frame. 
Melodic changes were annotated with WinPitch, allowing an easy graphic contour annotation from the existing segmentation into words and phones, as well as with an underlying narrow band spectrogram that would guarantee the validity of the annotated contour in term of frequency and duration (see Martin, 2018a). According to the incremental prosodic structure model (Martin, 2018b), prosodic annotations consist of melodic contours, mostly linear, located on stressed vowels (not syllables), and automatically classified from their slope in terms of direction and glissando values (i.e. perceived as a melodic change vs. perceived as a static tone): Cdec conclusive déclarative contour, perceived as sentence terminal; Cint conclusive interrogative contour, also perceived as sentence terminal; Cris rising continuation contour, as a melodic change; Cfal falling continuation contour, perceived as a melodic change; Cneu rising or falling contour, perceived as a static tone. Description of these contours in terms of frequency, duration and intensity, are automatically transferred on a spreadsheet program such as Excel for detailed phonetic analysis. In the intonation grammar of French retained for this research, the prosodic structure is defined by dependency relations “to the right”, i.e. when a given contour category depends on the category of another contour located further in the sentence. For instance, the falling contour Cfal depends on a rising contour Cris placed further in the sentence, Cris depends on a Cdec occurring later in the sentence, whereas Cneu depends on either a Cfal, Cris or Cdec / Cint located on its “right”. 
This intonation grammar is based on the contrast of melodic slope principle, essentially validated earlier on various read and spontaneous speech corpora involving mostly declarative sentences (see Martin, 2018b). The issue here is to validate this slope inversion principle for sentences using a final rising contour to indicate its interrogative modality. in which we would have Cfal depending on a terminal interrogative Cint, and possibly Cris depending on a Cfal located further in the sentence. However, most interrogative sentences of the corpus were rather of short duration, preventing the occurrence of a prosodic structure with 2 or 3 levels, therefore using the Cris -> Cfal dependency with the expected melodic contrast. Speaker sentences are usually quite short, with limited occurrences of reformulation and abandons, contrary to other spontaneous speech corpora. The speaker sentence prosody appears therefore rather well planned, striving for the shortest duration possible. Given the fact that clients had no clue in advance as to the personality of customer agent that will answer them, they usually start their calls with indirect questions with a declarative prosodic structure ended by Cdec, or alternatively with a statement describing the problem followed by a very short question associated to an interrogative prosodic structure (ended by a rising contour Cint). Likewise, if most customer use indirect questions in their demands, the eventual questions made by the agent to obtain some clarification use intonation to express an interrogative modality. 
In some recordings, agents will slow down the speech rate and expanding the melodic variations on stressed syllables in order to be better understood. Typical prosodic structure configuration such as long lists in route planning exchanges, segmentation of the sentence into prosodic nucleus and postnucleus (with a limited melodic excursion), etc. Analysis of abandons and reformulations in client sentences also illustrate the planning of the prosodic structure before the lexical items, which are substituted afterwards keeping the same melodic contour of the planned prosodic structure. 

References

Bechet F. et al. (2012) DECODA: a call-centre human-human spoken conversation corpus, Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC'12), pp. 1343-1347. 
Martin, Ph (2018a) Prosodic annotation in adverse recording conditions, a cura di A. De Dominicis, Speech audio archives: preservation, restoration, annotation, aimed at supporting the linguistic analysis, Accademia Nazionale dei Lincei, 
Martin Ph. (2018b) Intonation, structure prosodique et ondes cérébrales, London : ISTE, 322 p. 
ORFEO (2020) Corpus d’étude pour le français contemporain https://www.projet-orfeo.fr/ 
WinPitch (2020) Computer program, www.winpitch.com

Intonation of telephone conversations in a Customer Care Service

**Chair:** Fracnesca Dovetto (Università di Napoli Federico II) **Title:** EvaLatin 1.0: un Corpus per la Valutazione delle Tecnologie del Linguaggio Applicate al Latino 
**Author:** Rachele Sprugnoli, Matteo Pellegrini, Marco Passarotti, Flavio Massimo Cecchini
(Università Cattolica del Sacro Cuore, Milano; Università di Bergamo) 


 **Title:** Esplorazioni di treebank multilingue per studi tipologici 
**Authors:** Chiara Alzetta, Felice Dell'Orletta, Simonetta Montemagni, Giulia Venturi
(ILC-CNR; Università di Genova) 

**Title:** Intonation of telephone conversations in a Customer Care Service 
**Author:** Philippe Martin (Université Paris Diderot)

SESSIONE PLENARIA 2

keynote

### Benvenuti 
al *LIV Congresso internazionale della Società di Linguistica Italiana ''Corpora e Studi Linguistici''* che si tiene qui in rete nella piattaforma fornita da Underline. Il Congresso si sviluppa su 3 giornate divise in:

* cinque Sessioni plenarie, 
* due Sessioni Demo dedicate ai corpora italiani e internazionali, 
* una Sessione Poster e 
* una Sessione workshop con 5 workshop in parallelo. 

Gli interventi orali saranno sia in diretta che preregistrati, secondo la volontà degli autori, ma per tutti sarà possibile l’interazione dal vivo con domande e risposte. A tutte le sessioni si accede attraverso le icone sulla sinistra di questa pagina. 
 Nelle sessioni Poster e Demo ci si muove virtualmente tra le proposte attraverso Gather-Town, che permette ai congressisti anche di incontrarsi in qualsiasi momento nella Lounge virtuale.
 Ringraziamo tutti per l’impegno e la pazienza e ci auguriamo una buona riuscita del Congresso. Un ringraziamento particolare ai relatori invitati Bas Aarts (University College London), Gu Yueguo (The Chinese Academy of Social Sciences) e Mirko Tavoni (Università di Pisa) 
 

**Welcome** to the *54th Congress of the International Society of Italian Linguistics ''Corpora and Linguistic Studies''*
which is held here through the Underline Platform. The Congress develops in 3 days which host:

* five Plenary sections, 
* two Demo sessions, 
* dedicated to Italian and international corpus building initiatives, 
* one Poster session and 
* one Workshops session hosting 5 workshops in parallel. 

Oral presentations are given in synchronous or are pre registered according to the will of each author, however Q/A will be live for all. All sessions can be accessed through the icons on the left side of this page. 
 In the Poster and Demo sessions you will be able to move among the various items though Gather-town. Congress participants can always meet in the Virtual lounge.
 We wish to thank all participants for their work and patience. A special thanks to our invited speakers Bas Aarts (University College London), Gu Yueguo (The Chinese Academy of Social Sciences) e Mirko Tavoni (Università di Pisa)

![](https://assets.underline.io/uploads/markdown_image/1/image/dacffc533f05582ca945abd01bb616ea.png)

If you have already registered, please **login** to Underline.io to be able to see full conference content. 
 In case you do not have an Underline account, please **create** one. Thank you!

Philippe Martin

1

Presentations

Intonation of telephone conversations in a Customer Care Service