HK1237158B - Method and system for filtering undesirable incoming telephone calls - Google Patents

Description

Technical field

The invention relates to a process and system for filtering unwanted incoming telephone calls, including a process and system for detecting and preventing unwanted and criminal calls.

State of the art

Today's telephone communication systems make it easy to make unwanted calls, often with impunity. For example, many subscribers complain about receiving calls from advertising campaigns or opinion polls. Worse still, some calls may even be criminal, such as harassment, phone scam, fraud, or pedophile calls.

There are many solutions in the state of the art to block unwanted phone calls.

US2008292085A describes a device that blocks telephone calls from a blacklisted number. The caller has the option to add numbers to their list. However, this device provides no protection against unwanted calls from a non-blacklisted number, such as calls from a pay phone, or those whose caller number has been forged.

GB200614708A describes an inbound call filtering device that allows the user to program the numbers of authorized callers and block other calls. Users not on this list can enter a PIN code to be allowed to call. This device provides effective protection, but may create a significant number of false positives, i.e. calls that are blocked incorrectly even though they are not unwanted.

US2004131164 describes a device that allows calls to be routed to an answering machine when the caller ID is listed.

US8472599 describes a device that automatically rejects incoming calls when the caller ID cannot be determined.

US2002018546A describes a call filtering device that can be programmed to set different rules depending on the time of day or for different callers.

US2004086101A describes a telephone call filtering system suitable for use in, among other things, faxes.

EP0559047 describes a telephone call filtering device which can be used with a telephone receiver.

US8548149 describes a telephone answering machine capable of detecting unwanted calls, based in particular on patterns of calls and blacklists of unwanted callers.

WO11014103 describes a telecommunications system capable of filtering calls at the level of a telephone exchange.

JP2011010078 describes another junk filtering device that relies on blacklists of junk callers.

US2010278325A describes a method for predicting whether an incoming call will annoy the caller, and for preventing such a call.

EP2153637 describes a method for detecting unwanted phone call campaigns, for example by analysing the number of calls made from each number.

WO07144310A1 describes an incoming call management system that uses a blacklist, a whitelist, and a grey list of callers.

JP2007336001 describes a method to interrupt an unwanted call by gradually degrading the quality of the call.

WO07134810 describes a method for determining the probability that an incoming call is unwanted, based on the time of arrival of the call.

WO07118851 describes a method involving the creation of a blacklist of callers who have made at least one unwanted call.

US2009238345A describes a system for blocking unwanted calls before the first ring.

US2009238345A describes a device that allows a telephone number to be added to a blacklist or whitelist.

US5930700 describes a black and white list management system from a PDA.

US2013/100238 describes a system for routing calls based solely on video stream content.

Most of these known solutions are therefore based primarily on caller identification, or sometimes on other parameters such as the time of the call.

The current situation is that a person can easily call another person without being authorised, identified or even authenticated. For example, most operators allow the number of caller to be hidden. Anyone can also use a public telephone booth or simply make a VoIP call (e.g.: Skype) over the network to hide their identity.

As a result, the phone numbers that are displayed are sometimes fancy, or hidden, and do not really help callers make the right decision before they pick up.Some phone spammers even use caller numbers that correspond to government agencies whose calls must not, or cannot, be blocked under any circumstances.

While there is some state of the art in caller identification methods which are based on voice analysis rather than caller number, for example EP1564722 describes a process for routing and tracking telephone calls based on voice and caller number, which allows calls to be routed more efficiently to a call centre through improved caller recognition.

Caller voice recognition has also been used for unwanted call filtering. GB2474439 concerns a device for processing incoming phone calls to route them to the caller or reject them. The decision is made based on caller identification based on e.g. ID, PIN, voice recognition, or biometric information.

The identification by voice analysis must precede the establishment of the connection to the caller, and thus prolongs the phase of establishment of the call. For this reason, this relatively intrusive analysis is at most implemented in routing systems for call centers (as in EP1564722), since in these systems the caller is generally invited to speak to guide the routing anyway.

EP2306695 describes a process for identifying a caller in a telecommunications network, at the level of a call control node. If the caller is part of the known contact list, then the call is treated as a normal call and otherwise a second verification is carried out by sampling the caller's voice. This sample is then compared to the available voice samples and the call is rejected if there is no match and no connection between the caller and the caller. However, even if this comparison results in a match between the new sample and the available samples, this could possibly result in a mismatch.

Brief summary of the invention

One purpose of the present invention is to propose a process and method which will solve or at least reduce the problems described above.

In particular, one purpose of the present invention is to control incoming telephone access more effectively to effectively filter out unwanted calls.

A purpose of the present invention is to propose a process which is transparent and less intrusive for callers who are able and used to communicating with a given user.

According to the invention, these purposes are achieved in particular by means of a filtering process for unwanted incoming telephone calls, in which a decision to block or forward a call is based on both the caller number and an analysis of the caller's voice, the process involving the following steps: (a) determination of the caller number; (b) based on the caller number, a filtering system decides whether to forward the call, reject the call or invite the caller to speak; (c) when the caller is asked to speak, he or she speaks without having his or her voice message returned to the caller; (d) the filtering system analyses the caller's voice; (e) based on this analysis the filtering system decides whether the call should be forwarded to the caller, or blocked.

This method therefore uses at least two indicators, one on the caller number and the other which models the caller himself through voice analysis.

This method therefore makes it possible to block incoming calls that can almost certainly be classified as unwanted immediately; this can even be done without the ringing of the telephone being called.

Conversely, the procedure is transparent for calls which can be excluded from being telephone spam, particularly on the basis of the caller number.

In between, a special procedure is set up for calls that cannot easily be classified as unwanted or desirable; in this gray area, the caller is asked by the system to speak to the system in order to analyse his voice and determine, on the basis of that voice, whether the call should be forwarded or not.

This process prevents calls of a criminal nature even before they reach their victims.

The process may include a semantic analysis of the conversation, the decision to forward a call or the decision to block a call that depends on this semantic analysis.

In one embodiment, a score describes the trust the system places on a caller ID while a second measure verifies that it is the right person who is online by checking biometric fingerprint, e.g. voice, facial, multimodal (voice plus face), and possibly the caller's semantic fingerprint.

Caller number analysis may, for example, use blacklists, whitelists and/or grey lists of callers. These lists may be specific to each caller: a given caller may be undesirable to some callers but not to others. Other black, white and/or grey lists may also be shared between several callers, for example by storing them centrally in a shared server.

Synchronization mechanisms can be set up to synchronize shared lists with individual lists. In one embodiment, a caller is placed on a blacklist, respectively white or grey, and shared once a given number and/or percentage of callers have classified him as such.

Caller number analysis may also be based on caller number plausibility, for example to block numbers from non-existent or non-conforming callers (e.g. number of digits, area code, etc.).

Caller number analysis may also be based on the geographic origin associated with the forwarded caller number code; for a given caller, calls from abroad, or from a region or country other than those with which he or she is used to communicating, may be blocked or generate an invitation to speak before the call is forwarded.

Voice analysis of callers' voices allows, for example, to identify the person online and recognize them based on their voiceprint. Thus, a known telephone spammer (e.g. a call center employee) can be recognized based on their voiceprint even if they use different caller numbers. Voiceprints of unwanted callers can be shared between several callers.

Voice analysis also allows the caller to be authenticated and verified to match the transmitted caller number, when the fingerprint (s) associated with that caller number are known, for example on the basis of previous calls.

Voice analysis also allows the caller to be classified into a group among several groups, for example in a group blocked , admitted , etc.

A speech recognition algorithm and semantic analysis can be used to classify callers and to reject unwanted calls or accept desirable calls. For example, certain terms or phrases frequently used by intruders, or by a specific caller, can help trigger a block or acceptance decision.

The speech recognition algorithm can also compare the caller's speech with endogenous data (e.g. contents of previous dialogues with this caller known to the system) or exogenous data (e.g. data on a caller's or caller's social network).

The filtering system therefore uses caller identification through the voice of the caller. It can also use other information extracted from the received sound signal, including the detection of beep or sound signals emitted during calls from some call centers; background music detection, including the detection of particular tracks; background noise recognition and classification; automatic detection of synthetic voice; detection of already known audio extracts; semantic recognition of spoken text; etc. By combining several of these parameters, a probability or score of undesirable call detection can be determined.

According to the present invention, the filtering system continues to analyze the caller's voice after the call has been transmitted to the caller. This analysis allows, for example, to clarify the voice signature of the authorized callers whose call has been transmitted, in order to better recognize them in the future. It is also possible to block the call or propose to block it when this analysis reveals, after the call with the caller has been established, that it is undesirable, for example if it comes from a recognized intrusion late in the discussion, or when semantic analysis of his words reveals unacceptable content.

A confidence score may be assigned to the call based on the caller number, with the decision to reject the call or invite the caller to speak being made during step b when this score is below a predetermined threshold.

This trust score assigned to a caller is determined independently for each caller. For example, a caller number that is well known to a given caller, and has often called with him, will generate a higher trust score than a caller number that is unpublished for that caller. This individual trust score may also depend on the individual white, black, or grey lists specific to each caller as well as shared lists.

A second trust score can also be assigned to the call based on biometric analysis, e.g. voice, facial, voice plus facial, and/or semantic.

The trust score, and the decision to reject the call or invite the caller to speak may depend on the history of calls made from said caller number to other callers. For example, a caller number notoriously associated with nuisance calls may be mentioned in blacklists of numbers that can be shared, or conversely a government agency number may be listed on shared whitelists.

The process may include a caller identification step through biometric analysis, the trust score depending on the history of calls made by the caller thus identified to other callers. For example, a caller notoriously associated with nuisance calls may be listed in caller blacklists that can be shared.

The caller may prefer to report a call that is unwanted, for example via a button on a device, a voice command, or via a website. This may be done during the call, or after the call. The caller number may then be blacklisted, or associated with a lower trust score than the original. Similarly, the caller's biometric fingerprint, such as voice and/or facial, may be blacklisted, or associated with a lower trust score than the original. This decision may also be used in the decision to reject future calls from the caller to others, or to prevent the caller from speaking.

The process may employ a model of authorized callers, and another model of unauthorized callers. This model may be based on both the caller number and the caller biometric analysis. It may be individual for each caller. For example, an authorized caller model for a given call may contain the usual callers for that call, while an unauthorized caller model may contain callers on a blacklist as well as those whose caller number or voice analysis betrays a call from a call center abroad. Caller modeling may also be collective and identical for all callers.

The decision to accept a call and then forward it can be made when the parameters of that call are sufficiently close to the authorized caller model. The decision to block a call will be made when the parameters of that call are sufficiently close to the unauthorized caller model.

A machine learning algorithm can be set up to classify incoming calls into the category of authorised and unauthorised calls.

Brief description of the figures

Examples of implementation of the invention are given in the description illustrated by the figures in which:Figure 1 illustrates a telecommunications network including a filtering system in a first embodiment of the invention.Figure 2 illustrates a telecommunications network including a filtering system in a second embodiment of the invention.Figure 3 illustrates a telecommunications network including a filtering system in a third embodiment of the invention.Figure 4 illustrates a telecommunications network including a filtering system in a fourth embodiment of the invention.Figure 5 is a flow diagram illustrating an example of a process in accordance with the invention.

Examples of the implementation of the invention

Figures 1 to 4 illustrate four variants of filtering systems in a telecommunications network.

In the example in Figure 1, the telecommunications network consists of a telephone network 4, e.g. a POTS, ISDN, cellular, VoIP-type network, and/or a combination of several such networks. Reference 3 corresponds to a terminal device of a caller wishing to make a telephone call with a terminal device of a caller 2 through that telephone network.The functionality of module 1 can also be implemented by software. In one embodiment, this module is integrated into the terminal device of caller 2, for example in the form of firmware or an application, for example an application running on top of an operating system in the case of a terminal device in the form of a cell phone, tablet or computer in particular.The caller and caller number databases show, for example, which numbers and callers are on black, white or grey lists.

The filtering system 1.5 also includes a server 5 accessible from the telephone network 4 and containing a processor as well as a shared database of caller numbers 50, a shared database of callers 51 and a shared database of caller biometric fingerprints 52, e.g. voice and/or facial prints.

Server 5 may also contain call histories from different callers and/or to different callers, voice and/or biometric and/or semantic signatures of callers, etc.

Server 5 can also be realized as a virtual server or a group of interconnected servers, for example in the cloud.

The filter box 1 may be connected to this server 5 for example by means of an integrated modem 10 to establish a data communication with this server, for example during a non-dialogue interval or in a frequency band not used for voice communication.

The configuration in Figure 1 provides maximum installation comfort and avoids the call-taker the task of having to configure an internet connection; it is also appropriate when no internet connection is available. Typically, all landlines and mobile phones can thus be used by the invention. In this case only part of the information is communicated via the telephone line or mobile phone voice connection; typically the voice of the alleged perpetrator of an abusive call.

In the example in Figure 2, box 1 is connected to server 5 via the IP Internet network, for example via an Internet interface of box 1. The other elements can be identical to the corresponding elements in Figure 1. The IP connection expands the possibilities of use and improves the ergonomics of use. For example, it is possible to configure the filtering system via a web server, store parameters in server 5 (or in the cloud), and share information with other users.

In the example in Figure 3, the filtering system 5 is fully integrated into the telephone network 40, e.g. as a software and/or hardware module 5 which can be connected to a network node 40, e.g. an SSF node of a smart network.

In this mode of implementation, the service can also be carried out by diverting calls to the conventional telephone network or by diverting calls to another VoIP-type network. The user no longer needs to install a box at home and the interaction is done through web access or by a software application on the smartphone. The service can be managed by telephone operators or voice service operators, social networks, etc.

In the example in Figure 4, the caller's terminal device 2 is connected to server 5 via an Internet-like link, to allow the caller to, for example, report unwanted calls.

An example of a process used by the filtering system 1.5 of the invention is now described in relation to Figure 5.

During step a, a telephone call from telephone network 4 is received, and the caller line identity (CLI) is determined, e.g. in the device (or application) 1 if present, or in server 5 in the embodiments of Figures 3 and 4.

In step b, the caller number so determined is analysed, which may for example involve a comparison with a caller number list in a blacklist of unwanted numbers, and/or a comparison with a caller number list in a whitelist of unwanted numbers.

The comparison may be made with individual whitelists and/or blacklists to the caller, which may for example be stored locally in his device 1 or in a dedicated memory space of the server 5; these local lists may for example contain numbers that are unwanted (respectively acceptable) for a given caller but not necessarily for another caller. In addition, or alternatively, the comparison may be made with whitelists and/or blacklists which may be stored centrally in server 5 and which may contain unwanted (respectively acceptable) caller numbers for all or most of the callers. Synchronization mechanisms may be implemented for local lists and may be automatically synchronized in a centralized list, for example, for a predetermined number of callers, or a predetermined number of servers may be pre-incorporated in a predetermined number of servers.

The caller number analysis performed in step (b) may also include a caller number plausibility analysis, for example to exclude numbers whose format, number of digits or area code does not match a possible number, or corresponds to a number in a range of prohibited numbers.

The caller number analysis performed in step b may include a check of the history of calls made from that number, which are stored in server 5; thus a caller number associated with frequent calls over a short interval will have a higher probability of being classified as unwanted than a caller number used more conventionally.

The caller number analysis performed in step b can also include searching for this caller number in phone books, on the web, for example through search engines, in a social network like Facebook, Twitter, Linkedln etc, in a chat site etc and searching for a profile associated with this number in these different sources. These profiles, for example a company name, a business activity, a geographic location, are then used to calculate a score or probability of a call being unwanted. It is also possible to check whether a company or activity is part of a whitelist or a blacklist of companies or activities,For example, if it is a company known to generate phone spam. The score assigned to each company, activity or geographic location may be individual for each caller. A caller active in a given business area, or living in a given geographic location, is indeed more likely to receive calls from companies in or near that area than a caller active in another area or location. It is also possible to use a machine learning algorithm, based on a caller's past behavior, for example on accepted calls, their duration,on calls marked as unwanted, to calculate the individual score for that call.

A caller number associated with a company, or type of company, known for its phone harassment will thus be associated with a high probability of being unwanted. A particular caller may also decide to block, for example, all incoming calls whose number is associated in a directory or on the web to insurance companies.

Caller number analysis may also use an external and specialized certification entity, for example a PKI ( public-key infrastructure ) capable of certifying certain caller numbers

In step c (optional), a score is assigned to the call based on the results of the caller number analysis performed in step b. This score is related to the probability that the call will be unwanted for the given caller 2.

If the score is below a given first threshold, the incoming call is blocked (step m); optionally, a pre-recorded or synthetically generated voice message can be returned to caller 3 to inform him that the call is declined. If this score is above a given second threshold (above the first threshold), the call is passed to device 2 (step n) which starts ringing (or vibrating), so that the caller can answer. These secure calls are therefore established in a completely transparent manner for the caller and for the caller.

In the case where the score is between the first threshold and the second threshold, a special procedure is established, illustrated by the steps d to l in the figure.

The threshold levels 1 and 2 may optionally be adjusted by the user to adjust the false positive and false negative rates.

During step d, the caller is asked to speak, for example by means of a voice prompt synthesized by device 1 and server 5 respectively. The caller may be asked to enter his or her name and the name of the person desired. The invitation may be returned in such a way that the caller can hardly realize that he or she is speaking to a voice dialogue machine.

In one variant, the call is interrupted before step d. The system then establishes a communication in the other direction, i.e. by calling the caller back with the specified caller number, and asking them to speak. This variant is more constraining, but it provides additional security by allowing verification of the specified caller number.

During step e, the caller responds to the filtering system 1, 5, for example by giving his name and the person he wants.

During step f, the caller's voice is analyzed by the filtering system 1 and 5, respectively. The purpose of this analysis is to classify the caller as unwanted or not. A vocal extract of less than 30 seconds is usually sufficient.

In the case of multimodal communication, for example communication including image in addition to voice, it is optional to also analyse the caller's image and to perform facial analysis.

In the g-step (optional), a semantic analysis of the caller's voice is performed, for example on the basis of a voice-to-text conversion of the words spoken by the caller.The semantic imprint depends on all layers of language: speech, pragmatics, semantics, syntax, lexical level.It is for example possible to detect specific phrases, words or phrase turns that are more frequently found in unwanted calls.

During this step g, additional information can be extracted from the audio signal received from the caller. For example, many unwanted calls are generated from call centers that generate characteristic sound signals, such as beeps or other signals generated by the call center. Some operators in call centers work in music; sometimes a musical introduction is played as an introduction at the beginning of the call. Background noise, which is usually ruled out by voice or speaker recognition algorithms, can be very interesting because it allows to identify a central location, such as a calling location.For example, voice recognition systems can be specifically trained to recognize background noise in some call centers, and/or specific sound signals emitted by some call centers. The echoes emitted in the call room are also characteristic and depend on the geometry of the room. It is therefore possible to classify background noise and/or echoes during a call to determine whether it comes from a call center, or a specific call center.These extracts can be recognized. A synthetic voice detection module, e.g. one based on the measurement of the voice rhythm, can be used. These parameters, individually or in combination, constitute an acoustic fingerprint of the caller. They can be used to determine a score or probability of an unwanted call.

During the h (optional) step, the caller is modeled based on their voice and this semantic analysis. This step allows one to create a caller's voice signature, in order to recognize it in subsequent calls, and compare it to pre-existing models of unwanted users and acceptable users. This modeling may include, for example, a detection of the language spoken by the caller; it is thus possible to associate a language-dependent score. The model may also indicate emotional characteristics of the caller (e.g. stress state, aggressive tone, commercial score, etc.). The model may include an approximate determination of the caller's age and gender, using the k-scale and the probability of the caller being in the later stage of the call, and these parameters can be used to calculate the probability of the caller being in the same gender.

During step i, the caller is identified based on speech analysis.Identification may use speaker recognition algorithms and/or speech recognition algorithms to identify the caller based on his or her speech.

Identification is a determination of the identity of the caller, for example in the form of a unique identifier such as a name, number, etc. Identification may exploit the caller number previously determined during step a; however, several callers may share the same caller number, and a caller may use different caller numbers on different occasions.

Alternatively, or in addition, step i may also include caller authentication (or verification) to verify that the caller matches the specified caller number, the identity claimed in its voice response, and/or a group of claimed callers. For example, authentication may consist of verifying whether the caller's voice and/or facial signature belongs to the group of users who share a caller number or a group of caller numbers. In one option, the algorithm models in a combined manner (multimodal speaker recognition) using, for example, the consistency of the 3D representation with the visual appearances of spoken speech.

Alternatively, or in addition, step i may also include a classification of the caller among predetermined groups of callers. This classification may be based on, for example, models of predetermined caller groups. In one embodiment, at least one first model is established to model unwanted callers and at least one second model is established to model accepted callers. The classification may then involve calculating the distance between the caller model determined during step h and these two predetermined caller models, in order to classify the new caller into either class.

Classification may also involve classification into a group of unknowns, i.e. a group that includes all speakers outside the set of N caller identifiers known to the system.

Known machine learning technologies can be used for this classification (GMM, Bayesian network, Support Vector Machine), using both biometric and textual/semantic parameters.

Biometric modelling algorithms that are independent of spoken text and do not allow for the content of the text to be known can also be used, for example in a privacy-conscious approach and to avoid analysis of the content of conversations.

Data associated with the previously identified caller can be verified during the optional step j. For example, data associated with this caller and stored in device 1, server 5 or in external databases or websites can be exploited. In one embodiment, data relating to this caller is extracted from directories or social networks and can be used to assign a score to this caller during step l.

During the optional step k, data associated with the caller can be checked, for example data entered by the caller himself in the device 1, in the server 5, in a dedicated website, on a social networking site, a chat site, etc. This data may for example include lists of rejected callers, characteristics of rejected callers (e.g. based on language, age, gender, etc.), or points of interest or characteristics specific to the caller.

During step l, (optional), a score is assigned to the caller based on the results of steps f to k. This score is related to the probability that the caller is undesirable for the given caller 2. It can be combined with the caller number score determined during step c to calculate a combined score.

If the score is below a given third threshold, the incoming call is blocked (step m); optionally, a pre-recorded or synthetically generated voice message can be returned to caller 3 to inform him that the call is declined. If the score is above a given fourth threshold (above the first threshold), the call is forwarded to terminal 2 (step n) which starts ringing (or vibrating), so that the caller can respond.

When the call is established during step n, the caller and caller can converse normally. However, the caller can report at any time during step o that the caller is unwanted, for example by pressing a button on module 1, by a voice command, etc. In this case, the call is terminated and the caller number, as well as the caller's voice or biometric signature, are recorded as unwanted in the caller's personal blacklist.

During the dialogues between the caller and the caller, the voice and semantic analysis of the caller's words continues (steps f to l), and the score is continuously recalculated; the conversation can be automatically interrupted by the 1.5 system if this score falls below a given value. In one variant, the filtering system only suggests to the caller to interrupt the conversation, for example by means of a visual message displayed on a smartphone or on a module 1 screen, or a voice message rendered.

This analysis on the fly, during the conversation, also improves the caller's vocal and semantic pattern.

In the case of terminal device 2 consisting of a smartphone, tablet or computer, the confidence placed by the system in the call or in the authenticity of the conversation can be permanently displayed on the screen or by calling on the functionalities of the system used (vibrator, audible notification, etc.).

In the case of asymmetric smartphone-to-phone communication, the caller's biometric parameters may still include the visual modelling of the caller on the smartphone, and the parameters are then transmitted via another network connection to the filtering system 1, 5.

The system also offers the advantage of saving caller biometric data, e.g. voice and/or facial fingerprint. Criminal or unwanted callers leaving their biometric traces behind, this data can be shared with other members of the system user community, or with the police.

Now we'll describe an example of a dialogue in a system according to the invention. The system calls on a device that calls itself Katia. In this scenario, caller 3 is Alice and caller 2 is Bob.

In this example, the dialog is initiated by Alice dialing the number of the person called Bob. Bob has installed a filtering box Katia 1 between his terminal 2 and the telephone network 4, which intercepts this incoming call. Alternatively, Bob has installed an application on his smartphone that intercepts this incoming call.

System 1.5 detects that the score associated with the number is below the first trust threshold. In this case, the system picks up the call in place of Bob and performs a non-transparent access control for the caller. Specifically, box 1 picks up with the following message: Your number is not recognized by Katia, the rest of the conversation will be recorded, if you do not wish to be recorded hang up now. Katia is listening!

Such a message often already has the effect of deterring unwanted callers or criminals if they learn that their call is being recorded.

Alice responds with the sentence: "Hello Katia, I'm calling from New York. I need Bob's help, can you put him through?"

The system recognizes Alice through the stored voiceprint and also through the semantic fingerprint created by the Facebook page where Alice had described her trip to New York. With a favorable result of the probability calculation (the call number score combined with the probabilities that result from the fingerprint verification), the system grants access and rings Bob's phone.

In a preferential mode, Bob has listened to Alice's message being returned to him, allowing for manual filtering by Bob. Bob can communicate to device 1 his decision to take the call e.g. with a voice command of type Katia, I take the call . Communication is then established between Alice and Bob.

Alice then speaks to Bob. In this example, she asks: "Hello Bob, can you send me money to my account? It's urgent!" . System 1.5 records the conversation and segments the speech of each speaker. Alice's semantic footprint is continuously analyzed to check for spam. Bob answers and the conversation continues.

The filtering system 1.5 updates the voice, biometric and semantic patterns of the caller Alice.

The present invention also relates to a tangible computer data carrier containing a program which can be run by a computer system 1.5 to perform all or part of the above process.

Claims (18)

1. A process for filtering unwanted incoming telephone calls, in which a decision to block a call or to transmit it is based both on the calling party number and on an analysis of the voice of the calling party, the process having the following steps:

   a) determination of the calling party number;

   b) on the basis of the calling party number, a filtering system (1, 5) decides either to transmit the call, or to reject the call, or to ask the calling party (3) to speak;

   e) when the calling party (3) is asked to speak, he speaks without his voice message being reproduced for the calling party;

   f) the filtering system analyzes the voice of the calling party (3);

   I) the filtering system takes this analysis as a basis for deciding whether the call needs to be transmitted to the called party (2), or whether it needs to be blocked,

   characterized in that the filtering system continues to analyze the voice of the calling party (3) after the call has been transmitted to the called party (2) and blocks the call or proposes to block it when this analysis shows, after the call has been transmitted to the called party (2) that the call is unwanted.
2. The process according to claim 1, in which the decision to block a call or to transmit it in the course of step I) is moreover based on an analysis of the face of the calling party (3).
3. The process according to one of the claims 1 to 2, in which said filtering system can, in the course of step b, also decide to call back the calling party (3) in order to check his number and to ask him to speak.
4. The process according to one of the claims 1 to 3, comprising a step (c) in the course of which a confidence score is attributed on the basis of the calling party number, the decision to reject the call or to ask the calling party (3) to speak being taken in the course of step b) when this score is below a predetermined threshold.
5. The process according to claim 4, in which said confidence score attributed to a calling party (3) is determined independently for each called party.
6. The process according to one of the claims 4 or 5, in which said confidence score, or another confidence score, is attributed to each calling party (3) on the basis of said voice analysis.
7. The process according to claims 4 to 6, in which said decision to reject the call or to ask the calling party (3) to speak is dependent on the history of calls made with said calling party number to other called parties.
8. The process according to claims 4 to 7, comprising a step (i) of identification of said calling party (3) by courtesy of said voice analysis, the confidence score being dependent on the history of calls made by said calling party (3) identified in this manner to other called parties.
9. The process according to claims 4 to 8, comprising a step (i) of authentication of said calling party (3) courtesy of said voice analysis and of checking that said authentication corresponds to the calling party number.
10. The process according to claims 4 to 9, comprising a step (j ; k) of checking a calling party profile in a directory, on the web and/or in a social network, said decision to reject the call or to ask the calling party (3) to speak being dependent on these data.
11. The process according to claims 1 and 4 to 10, comprising a step (o) in the course of which the called party (2) signals an unwanted call, this indication then being used in the decision to reject future calls from the calling party (3) to other called parties, or to ask this calling party (3) to speak.
12. The process according to claims 1 and 4 to 11, comprising a step (g) of semantic analysis of the conversation, the decision to transmit a call or the decision to block a call being dependent on this semantic analysis.
13. The process according to claim 12, the decision to transmit a call or the decision to block a call on the basis of the semantic analysis being dependent on each called party (12).
14. The process according to one of the claims 1 and 4 to 11, comprising a step (g) of analysis of the background noise, of the echoes, of the music and/or of the sound signals transmitted by the calling party (3), the decision to transmit a call or the decision to block a call being dependent on this analysis.
15. The process according to claims 1 and 4 to 14, comprising a step (h) of modelling of the authorized calling parties, along with a step of modelling of the unauthorized calling parties.
16. The process according to claims 1 and 4 to 15, in which said decision to transmit the call, to reject the call, or to ask the calling party (3) to speak is based on a machine learning method.
17. The process according to claims 1 to 16, in which said filtering system (1) performs a plausibility analysis for the calling party number and takes this plausibility analysis into account in its decision either to transmit the call, or to reject the call, or to ask the calling party (3) to speak.
18. A process for filtering unwanted incoming telephone calls comprising:

    a module for determining the calling party number;

    a filtering system (1, 5) capable of deciding, on the basis of the calling party number, whether the call needs to be transmitted, rejected, or whether the calling party (3) needs to be asked to speak;

    a module for analyzing the voice of the calling party (3);

    the filtering system moreover being designed to take this analysis as a basis for deciding whether the call needs to be transmitted to the called party, or whether it needs to be blocked, characterized in that the filtering system is moreover designed to continue to analyze the calling party's (3) voice after the call has been transmitted to the called party (2), and to block the call or to propose to block it when this analysis shows, after establishing the call with the called party, that the call is unwanted.