Several phenomena of disfluency can be observed in the utterances of spontaneous oral conversation. They involve the phonetic, acoustic and prosodic levels as well as the morpho-syntactic level. This study focuses on the morpho-syntactic level. Variations in the rhythm of verbal fluency (Pasdeloup, 1992; Duez, 2007; Béchet et al. 2013) can extend to the self-interruption of words or phrases These self-breaks and auto-variations in the verbal fluency are related, in most of the cases, with one or several kinds of events or items inserted in the middle of a phrase or even a word. They can cause disturbances in the morpho-syntactic organization of verbal flow, the most frequently quoted being the resumptions after a break, such as auto-repairs and incomplete phrases or words (Clark & Wasow, 1998; Henry & Pallaud, 2003; 2007; Guenot, 2005; Pallaud & Henry, 2006b).Rather than trying to quantify and analyze some phenomena of disfluency previously defined (such as incomplete words or repetitions, for example), the preferred option was to identify all the auto-breaks in the statements and to describe the morpho-syntactic consequences of these interruptions in verbal flow. This approach, contrary to some previous work (for example Constant & Dister, 2010), led us to distinguish what a self-interruption index is, i.e., the frequent signs of auto-breaks (silent or filled pauses, discursive markers, for example), and the effects of these ruptures on verbal flow. This made it possible on the one hand to exhaustively identify all the auto-breaks in a statement and on the other hand to describe the relations existing between these auto-breaks, their signs and their morpho-syntatic effects, which results in better specifying the concept of disfluency. In order to describe all of the breaks in the statements of four corpora of the CID (Corpus of Interactional Data) and all of the morph-osyntactic disturbances, we successively used two methods of detection, one semi-automatic (detection of all Interregnum spaces; Shriberg 1994) and the other, a manual one. Both are used within the framework of the Praat software as a tool for identifying, describing and annotating the phenomena (Blache et al. 2010). The automatic method made it possible to identify only 81% of the breaks, the 19% remainder, were manually identified as being disfluent.Main resultsThe breaks, thus stated, are numerous. On average, it is possible to find one rupture in the syntagmatic flow every 7.4 words (from 6.2 to 9.8 words, depending on the speakers). However, when the syntagmatic flow is stopped, it is not always broken: half of these ruptures are just hung up i.e. the statement is going on as if it had not been suspended. The other half causes a morpho-syntatic disturbance (unfinished or resumed statements); also their frequency strongly varies from one speaker to another: on average, it is one every 15.9 words.Therefore, half of the detected breaks are on average, just suspending ones; the other half are disfluent. Whoever the speaker may be, among all the disfluent breaks, those which are followed by a resumption of the statement are far more numerous (62% versus 38%) than those where the statement is left unfinished (results already obtained during recorded semi-directed talks (Pallaud & Henry, 2007). The resumption of the statement after a disfluent auto-break seems thus to be a main characteristic in the oral statements.This study equally analyzes the relationship existing between the suspending or disfluent breaks and the presence of a sign of interruption. If, in 81% of the cases only, the auto-interruptions, whatever they are, are announced by signs (silent or filled pauses, discursive interjections, markers), the others (19%) were detected only by reading/listening which allows for the identification of resumptions or abandoning of a phrase or a word as they are not related to signs. This last category of breaks is by definition disfluent since, no signs being related to it, the disturbance of the statement (resuming or rupture patterns) is the only phenomena that makes it possible to detect these disfluencies. This fact has implications for the development of methods for automatic control of disfluent breaks.All the types of insertion occur in both categories of breaks (suspendeding or disfluent ones) but in different proportions. We distinguished, in the Interregnum spaces, the silent or filled pauses alone, the speech markers alone and multiple spaces (containing more than one of the previous elements). The indicators related to the auto-break are not the same if the break is just suspending or if it is disfluent. When the Interregnum space contains a silent pause alone or a speech marker alone, in two cases out of three, this space follows a simple suspending of the statement. Moreover, the category of the disfluent breaks is the only one where the Interregnum space can be empty; it is the case for 39% of these breaks (which, therefore, were not detected automatically). In addition to their morpho-syntactic effects, the two kinds of disfluent interruptions are distinguished by the content of the Interregnum space which succeed them. Only a quarter of the disfluencies left unfinished are followed by a filled Interregnum space whereas half of the disfluencies followed by a resumption are. Thus, the speakers back up in their statement and restart immediately in half of the cases whereas that occurs only in one quarter of the cases of a phrase or clause left unfinished. More precisely, when the Interregnum space is filled, the silent pauses and the speech markers (each alone) are in an equivalent proportion in the case of a restart or of an abandoning. It is not the same when the Interregnum space is filled with filled pauses or multiple spaces. The disfluencies left unfinished are characterized by filled pauses alone or multiple spaces two times more frequent than in the disfluencies followed by a resumption,The dialogs in the CID were organized in order to allow the analysis of the statements uttered by the two speakers during their conversation, which lasted one hour. This kind of interaction is characterized by a succession of talks that may overlap from time to time. On average, the duration of the whole talk for each speaker is 33 min and varies little according to the speakers (from 31.7 to 35;4 min). The evaluation of this duration comprises the duration of all the informative utterances (ideal statement) and the duration of the disfluent elements (Interregnum spaces and “repaired” items). We found that the percentage of thus defined disfluencies compared to the contents of the oral emissions varies from one speaker to another: it is 38% on average (33,3%< X <40,1%). As Bear and al. (1992) mentioned, “In summary, disfluencies occur at high enough rates in human-computer dialog to merit consideration.” And that is not limited to the human-machine dialogs. This corpus of oral spontaneous statements shows that conversations also contain a lot of interruptions and several sorts of disfluencies.