Processing method implemented by a virtual assistant system and corresponding system

Abstract

The present invention relates to a processing method implemented by a virtual assistant system (GA1) and a corresponding system. The processing method comprises: detecting, from audio data (DT1), audio segments (SG1) each containing a speech activity (8) of a speaker (UR1); assigning a speaker identifier (ID1) to each audio segment (SG1); determining transcription data (DT2) associating a transcription (DT3) of the audio segments (SG1) with the speaker identifiers (ID1); and sending, to a massive language model (ML2), an instruction (RQ1) to trigger the execution of a function (FN1), said instruction specifying said function (FN1) in association with the transcription data (DT2) serving as context for performing said function. Figure 1

Description

Title of the invention: Processing method implemented by a virtual assistant system and corresponding system. Technical field

[0001] The present invention relates to the fields of audio data processing and text recognition from audio data. More particularly, the invention relates to a virtual assistant system and a processing method implemented by such a system to perform a function from audio data. Technological background

[0002] Virtual assistant systems, also known as virtual assistants or chatbots, have profoundly transformed our interaction with technology in recent years. Made possible by major advances in artificial intelligence and natural language processing, these systems have opened up new perspectives in areas such as human-computer interaction, information retrieval, and the automation of everyday tasks. In this context, virtual assistants have become important tools in both professional and personal life, offering rapid access to a multitude of services and information, particularly in the field of audio processing.

[0003] The operation of virtual assistants relies in particular on machine learning and speech recognition algorithms. These systems are capable of processing complex requests and providing relevant answers using sophisticated language models. Interaction with a virtual assistant can take place through various channels depending on the situation, including speech, text, and / or graphical interfaces, which makes them particularly versatile and accessible.

[0004] However, despite the opportunities offered, virtual assistants still face significant limitations and constraints.

[0005] For example, the "Intelligent Recap" feature offered in Microsoft® Teams Premium® is an automatic online meeting summary tool that uses artificial intelligence (AI) techniques. This service works by analyzing all the content of a meeting, including discussions, shared documents, and key points addressed. A summary highlighting key moments, action items, and essential meeting information can then be automatically generated. This solution offers significant time and efficiency gains, as users can stay informed about the key points discussed during a meeting without having to listen or search. the entire recording of the meeting.

[0006] However, virtual assistants, such as those implementing the aforementioned "Intelligent Recap" function or a similar function, have limitations, particularly in their ability to perform tasks using audio data. Current virtual assistants, for example, are capable of processing audio data from a call and differentiating between different speakers participating in the call (diarization process). In contrast, different speakers during an audio conversation (during a telephone call or video conference, for example) are identified as distinct speakers only if they log in separately to the audio conversation session.If several individuals participate in the conversation via the same customer account (for example, using the same communication device in a meeting room), then all of these individuals are treated as a single speaker, which limits the service capabilities of the virtual assistant.

[0007] In addition, an "Intelligent Recap" type function or equivalent requires recording the audio conversation and subsequent processing of the recorded audio data to generate an output meeting summary, which prevents any real-time processing during an audio conversation, and therefore does not allow certain services to be provided immediately to users.

[0008] More generally, current virtual assistants are limited in their ability to perform various functions from audio data, for example, to provide services during audio sessions or conversations (calls, video conferences, etc.). Summary of the present invention

[0009] One of the objects of the present invention is to solve at least one of the problems or deficiencies described above.

[0010] Another object of the present invention is to enable the reliable and efficient implementation of a function or service from audio data.

[0011] Another object of the present invention is to enable enriched and improved processing of audio data representative of the vocal activity of multiple speakers.

[0012] According to a first aspect, the present invention relates to a processing method implemented by a virtual assistant system, said method comprising: - detection, from audio data, of audio segments each comprising a speaker's vocal activity; - assigning a speaker identifier to each audio segment respectively; - determination of transcription data associating respectively a transcription of the audio segments with speaker identifiers; and - sending, to a massive language model, an instruction to trigger the execution of a function, said instruction specifying said function in association with the transcription data serving as context to perform said function.

[0013] The present invention advantageously allows for the identification of speakers in audio data representative of a speech exchange, without the need for external information about the speakers (no need for an external audio source, audio sample, or predefined personal information). The identity of the respective speaker is included in the transcription data in association with each audio segment transcription, which allows the massive language model to differentiate the audio segment transcriptions according to their respective speaker. The addition of speaker identifiers to the transcription data enriches the context provided to the massive language model, thus enabling it to perform the function more reliably and efficiently. A massive language model is not capable of distinguishing different speakers without these speaker identifiers.

[0014] Thanks to the invention, tasks normally performed by people, for example employees or others, can be advantageously automated, thus achieving significant savings in time and human resources, while requiring moderate processing resources. For example, the invention makes it possible to automate note-taking during meetings. Minutes can be generated automatically and distributed to interested parties with significant gains in time and efficiency.

[0015] The method according to the invention may include other features which may be taken separately or in combination, in particular among the following embodiments.

[0016] According to a particular embodiment, the transcription data specified in the instruction define a context in which the function must be performed by the massive language model.

[0017] According to a particular embodiment, the method includes, prior to the detection of audio segments, obtaining the audio data by extraction from at least one audio file or from at least one audio stream.

[0018] According to a particular embodiment, a diarization process is used for the detection of audio segments and the assignment of speaker identifiers.

[0019] According to a particular embodiment, the determination of transcription data includes: - transcription of audio segments into text data from a voice-to-text transcription model; and - encoding of text data in association with the respective speaker identifiers in the transcription data.

[0020] According to a particular embodiment, the function specified in the instruction comprises: - generation of a summary of the voice activity contained in the audio data for a given speaker.

[0021] According to a particular embodiment, the method comprises obtaining a number of speakers participating in the audio data, in which the assignment of a speaker identifier to audio segments is done from the number of speakers.

[0022] According to a second aspect, the present invention relates to a device configured to implement a virtual assistant system according to the first aspect of the invention. The invention particularly relates to a device configured to implement a virtual assistant system in cooperation with a massive language model, said device comprising a memory associated with a processor configured to implement the processing steps according to the first aspect of the present invention.

[0023] It should be noted that the various embodiments mentioned above in relation to the processing method according to the first aspect of the invention and the associated advantages apply in a similar way to the virtual assistant system according to the second aspect of the invention.

[0024] According to a third aspect, the present invention relates to a processing system comprising a virtual assistant system according to the second aspect of the present invention and a massive language model configured to perform a function in response to an instruction sent by the virtual assistant system. This virtual assistant system can, for example, be implemented in a motorized land vehicle, such as a car or other vehicle.

[0025] According to a fourth aspect, the present invention relates to a computer program comprising instructions adapted for executing the steps of the processing method according to the first aspect of the present invention, particularly when the computer program is executed by at least one processor. In other words, the various steps of the processing method are determined by computer program instructions. This computer program is configured to be implemented in a virtual assistant system of the second aspect of the invention, or more generally in a computer.

[0026] Such a computer program may use any programming language, and be in the form of source code, object code, or an intermediate code between source code and object code, such as in a partially compiled form, or in any other desirable form.

[0027] According to a fifth aspect, the present invention relates to a recording medium (or information medium), readable by the virtual assistant system according to the second aspect or more generally by a computer (or a processor), on which is recorded a computer program comprising instructions for the execution of the steps of the treatment process according to the first aspect of the present invention.

[0028] On the one hand, the recording medium can be any entity or device capable of storing the program. For example, the medium can include a storage means, such as a ROM, a CD-ROM or a microelectronic circuit-type ROM, or a magnetic recording means or a hard disk drive.

[0029] On the other hand, this recording medium can also be a transmissible medium such as an electrical or optical signal, such a signal being able to be transmitted via an electrical or optical cable, by conventional or radio frequency, by self-directing laser beam, or by other means. The computer program according to the present invention can, in particular, be downloaded from an Internet-type network.

[0030] Alternatively, the recording medium may be an integrated circuit in which the computer program is incorporated, the integrated circuit being adapted to execute or to be used in the execution of the process in question. Brief description of the figures

[0031] Other features and advantages of the present invention will become apparent from the description of the particular and non-limiting embodiments of the present invention below, with reference to the attached Figures 1 and 2, in which:

[0032] [Fig-1] schematically illustrates a virtual assistant system cooperating with a massive language model within a processing system, according to at least one particular and non-limiting embodiment of the present invention; and

[0033] [Fig.2] illustrates a diagram of different stages of a processing method implemented by the virtual assistant system of [Fig.1], according to at least one particular and non-limiting embodiment of the present invention. Description of examples of achievements

[0034] A processing method and a virtual assistant system configured to implement such a method will now be described in the following according to particular embodiments of the invention with joint reference to Figures 1-2. Unless otherwise indicated, common or similar elements in several figures bear the same reference numerals and have identical or similar characteristics, so that these common elements are generally not described again for the sake of simplicity.

[0035] The terms "first(s)" (or first(s)), "second(s)", etc.) are used in this document by arbitrary convention to allow identification and distinction of different elements (such as operations, means, etc.) implemented in the embodiments described below.

[0036] As previously stated, the invention relates in particular to a processing method implemented by a virtual assistant system. The terms "virtual assistant system," "virtual assistant," and "conversational agents" may be used interchangeably in this disclosure. A virtual assistant is a system, implementing a computer program, configured to perform operations or services automatically in order to assist one or more users in certain tasks. To do this, the virtual assistant can take as input certain data or commands, in oral, written, or other form, so as to perform a given function as output.

[0037] The processing method of the invention relies in particular on the ability of a virtual assistant to distinguish different speakers in an audio conversation and to use a massive language model to perform a given function. To do this, the virtual assistant can identify different speakers in audio segments of an audio recording and transmit to a massive language model transcription data comprising speaker identifiers associated with the transcription of the respective audio segments.

[0038] A large language model, also called a massive language model, large linguistic model, or LLM, is a type of artificial intelligence (or predictive) model configured to analyze and process text data. These models are generally based on deep neural networks that are trained on textual data to learn and recognize written language. The term "massive" refers to the large number of parameters that this type of model contains, which allows it to accurately capture the nuances of a language.

[0039] According to particular embodiments, the processing method comprises: - detection, from audio data, of audio segments each comprising a vocal activity of a speaker; - assigning a speaker identifier to each audio segment respectively; - determination of transcription data associating respectively a transcription of the audio segments with speaker identifiers; and - sending, to a massive language model, an instruction to trigger the execution of a function, said instruction specifying said function in association with the transcription data serving as context to perform said function.

[0040] This method advantageously enables the reliable and efficient implementation of a function using audio data representative of an audio conversation or vocal activity. In particular, it makes it possible to improve the processing of audio data representative of the vocal activity of multiple speakers for the purpose of performing a function (or service), such as, for example, the automatic generation of a summary of meeting or any other automated task that can be performed by a massive language model from such audio data.

[0041] Other aspects and advantages of the present invention will become apparent from the embodiments described below with reference to the drawings mentioned above.

[0042] Figure 1 schematically illustrates a first device DV1 comprising a virtual assistant (or virtual agent) AGI system configured to cooperate with a massive language model ML2. In this example, this massive language model ML2 is implemented by a second device DV2 distinct from the first device DV1, although variations are possible. More generally, the devices DV1 and DV2 can be distinct or form a single device.

[0043] As illustrated in [Fig.1], the devices DV1 and DV2 form in this example a processing system SY1. These devices DV1 and DV2 can include all technical means (hardware and software) necessary for the execution of the virtual assistant AGI and the massive language model ML2 respectively.

[0044] The AGI virtual assistant is configured to obtain DTI audio data representative of the voice activity 8 of a plurality of speakers URL. These speakers UR1 are participants, or users, who speak (or emit sounds) during a conversation or audio session. This voice activity 8 can be captured in various ways and can take various forms depending on the case. For example, it could be a conversation or voice exchange between speakers UR1 during a meeting, a telephone call, or a videoconference, or any other communication session involving the generation of voice activity 8 by the speakers in question.

[0045] The way in which the AGI virtual assistant obtains the DTI audio data may vary. The AGI virtual assistant may, for example, use sound sensors (not shown), such as microphones or other devices, to determine the DTI audio data. Alternatively, the AGI virtual assistant may retrieve the DTI audio data from a memory location (for example, memory 16) in which said DTI data is stored.

[0046] In the example considered, the DV 1 device is a processing device comprising at least one processor 10, a volatile memory 12 and a non-volatile memory 14. The DV 1 device may, for example, comprise all or part of a computer or server, or a combination of computers or servers.

[0047] To this end, the DV 1 device may include a computer program PG1 stored in the non-volatile memory 14 accessible by the processor 10, this computer program PG1 comprising instructions for carrying out a processing method according to a particular example of the invention. The processor 10 is thus configured to execute, in particular, the instructions defined by the program computer program PG1. In this example, the DV1 device is configured to implement the AGI virtual assistant by running the computer program PG1.

[0048] The non-volatile memory 14 can constitute an information carrier according to a particular embodiment in that it includes the computer program PG1 comprising instructions for carrying out the steps of a processing method of the invention which is described later in particular examples.

[0049] The DV1 device may further include a second memory 16 accessible by the processor 10 and suitable for storing data processed or generated during the execution of the processing method according to the computer program PGL. This second memory 16 may be all or part of at least one of the memories 12 and 14 or be a separate memory accessible by the processor 10. More particularly, the memory 16 may be used to store the DTI audio data obtained by the AGI virtual assistant as well as DT2 transcription data generated by the AGI virtual assistant during the processing method, as described later.

[0050] The AGI virtual assistant is configured to perform processing on DTI audio data. As described below in specific examples, this processing includes diarization and transcription processes on the DTI audio data. Diarization is a processing step that segments the DTI audio data into different SGI audio segments (or parts), identifies the individual speakers UR1 present in the DTI audio data, and assigns a respective speaker identifier ID1 to each SGI audio segment. Each ID1 corresponds to a respective speaker UR1. The DTI audio recording can thus be divided into distinct SGI segments associated with each individual UR1 who speaks.

[0051] As described below, the AGI virtual assistant can implement modules to perform the diarization process of UR1 speakers in DTI audio data.

[0052] The AGI virtual assistant can also be configured to use an ML1 transcription model to perform the transcription process of the SGI audio segments contained in the DTI audio data. To do this, the ML1 transcription model can be implemented using a neural network of encoder-decoder transformers, as described below.

[0053] As illustrated in [Fig. 1], the AGI virtual assistant is further configured to cooperate with the ML2 massive language model, in particular to send an instruction (or command, or "prompt") RQ1 aimed at triggering the execution of a function (or service) FN1 by the ML2 massive language model. To this end, a suitable communication link can be established between the DV1 and DV2 devices. Variants are also possible in which the DV1 and DV2 devices are part of the same device.

[0054] As illustrated in [Fig. 1], the second device DV2 may include at least one processor 20 configured to execute the massive language model ML2 stored in a memory (not shown). The massive language model ML2 includes at least one neural network (called a deep neural network) configured to process DT2 transcription data received as input and to perform as output a function denoted FN1, which may be of various kinds depending on the case, as described below in specific examples.

[0055] The FN1 function performed by the ML2 model can vary depending on the case. It can be any appropriate function that one wishes to perform from DT2 transcription data representative of the vocal activity of the URL speakers.

[0056] According to one example, the massive language model ML2 is specialized to perform specific tasks defined by the function(s) FN1, such as an automatic translation operation, answering questions, automatically generating a summary, etc.

[0057] The training method implemented by the AGI virtual assistant of [Fig. 1] is now described with reference to [Fig. 1] according to particular embodiments. To this end, the DV1 device implements the AGI virtual assistant by executing the PG1 computer program ([Fig. 1]).

[0058] In what follows, we consider, by way of example, that users UR1 produce respective voice activities 8, for example during a meeting (in person, by telephone, by videoconference, etc.). These voice activities 8, including spoken language, are captured as DTI audio data (or voice data) which are obtained by the AGI virtual assistant. In other words, the AGI virtual assistant obtains (determines or retrieves) DTI audio data representative of the voice activity 8 of the UR1 speakers conferring together. This DTI audio data can be captured in real time during a voice exchange between the UR1 speakers. Alternatively, the DTI audio data is stored in the memory 16 of the DV1 device and retrieved by the AGI virtual assistant from memory 16.

[0059] According to one example, the DTI audio data is retrieved by the AGI virtual assistant in the form of an audio stream. Therefore, it is not necessary to record the DTI audio data.

[0060] According to one example, the AGI virtual assistant obtains DTI audio data by extraction from at least one audio file and / or from at least one audio stream. The form and content of this DTI audio data may vary depending on the case.

[0061] During a detection step S2 ([Fig.2]), the AGI virtual assistant detects, from the DTI audio data, SGI audio segments, each comprising a respective voice activity 8 of a speaker UR1 among the speakers considered. The AGI virtual assistant defines, for example, a start time and a duration for each SGI audio segment.

[0062] During an assignment step S4 ([Fig. 2]), the AGI virtual assistant assigns a speaker identifier ID1 to each detected SGI audio segment in the DTI audio data. By assigning a speaker identifier ID1 to each SGI audio segment, each SGI audio segment can be associated with a corresponding speaker UR1 who produces the speech activity 8 contained in said SGI audio segment. For example, the speech activity 8 of a single speaker UR1 is contained in each SGI audio segment.

[0063] By way of example, the AGI virtual assistant can detect (S2) 4 audio segments SG1, SG1, SG1c and SG1d and assign (S4) to them speaker identifiers ID1, ID1, ID1d and ID1d, respectively. These ID1 identifiers can, for example, include at least one alphanumeric or other character.

[0064] The S2 detection of the SGI audio segments and the S4 assignment of the ID1 speaker identifiers can be carried out during a process called "diarization" which allows the differentiation of the different vocal activities 8 contained in the DTI audio data according to the speakers UR1 at the origin of said vocal activities 8.

[0065] According to one example, the AGI virtual assistant comprises modules cooperating to perform the speaker diarization process UR1 (steps S2 and S4), namely a speech recognition module to identify speech activities 8 (see, for example, Gelly et al., “Optimization of RNN-Based Speech Activity Detection,” IEEE / ACM Transactions on Audio, Speech, and Language Processing, vol. 26, no. 3, pp. 646–656, March 2018), a detection module to detect a speaker change (see, for example, Yin et al., “Speaker Change Detection in Broadcast TV Using Bidirectional Long Short-Term Memory Networks,” Proc. Interspeech, 2017), and an overlay module to process an overlay of several speech activities (see, for example, Bullock et al., “Overlap-Aware Resegmentation for Speaker Diarization,” submitted to ICASSP). 2020). A grouping module can also be used to identify the same UR1 speaker across multiple SGI segments.These modules can be combined to form a coherent speaker diarization process, or “pipeline” (see, for example, Bredin et al., “py annote, audio: neural building blocks for speaker diarization”. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 2020).

[0066] During a determination step S6 ([Fig. 2]), the AGI virtual assistant determines DT2 transcription data, associating a DT3 transcription of the SGI audio segments with the speaker identifiers ID1. Thus, the DT2 transcription data includes a DT3 transcription of each SGI audio segment in association with the corresponding speaker identifier ID1. These DT3 transcriptions are textual representations of the vocal activities (or speech) recognized in the SGI audio segments. These DT3 transcripts are, for example, text data defining textual representations of the vocal activities recognized in the SGI audio segments. The DT3 transcripts (or text data) can thus be encoded in the DT2 transcription data in association with the respective ID1 identifiers.

[0067] According to one example, the AGI virtual assistant can be configured to use an ML1 transcription module ([Fig. 1]) to determine the DT2 transcription data during the S6 determination step. This ML1 model makes it possible, in particular, to transcribe the SGI audio segments contained in the DTI audio data into text (speech-to-text conversion) and to associate the textual transcription of each SGI audio segment with a respective speaker identifier ID1.

[0068] The S6 determination of DT2 transcription data includes, for example: - the transcription of SGI audio segments into DT3 text data; and - the encoding of DT3 text data in association with the respective ID1 speaker identifiers in the DT2 transcription data.

[0069] The AGI virtual assistant is thus advantageously able to identify in the DT2 transcription data which speaker UR1 said or vocalized each SGI audio segment, respectively. Each SGI audio segment can thus be individually transcribed (or converted) into text while preserving the information relating to the identity of the associated speaker UR1.

[0070] According to one example, the AGI virtual assistant includes a neural network of encoder-decoder transformers to implement the ML1 transcription model and thus determine the DT2 transcription data during the S6 determination step (cf. for example Vaswani et al. “Attention is all you need”, Advances in neural information processing Systems, pp. 5998-6008, 2017).

[0071] In particular, the ML1 transcription model ([Fig. 1]) is a sequence-by-sequence prediction model based on the attention mechanism. This model is based on a transformer architecture, rather than recurrent neural networks (RNNs). These transformers (also called transformers, or self-attentive models, or simply "transformers") are configured to handle sequential data, such as language, to perform tasks. However, unlike recurrent neural networks, transformers do not require that sequential data be processed in order. For example, if the input data forms a sentence in a given language, a transformer does not need to process the beginning before the end. The transformer allows for much greater parallelization than recurrent neural networks and therefore reduced training times.

[0072] The transformers rely on the attention mechanism without using a recurrent neural network structure. This attention mechanism allows the ML1 model to The attention layer can directly examine and extract the state of any previous token in the sequence. It can then access all previous states and assign them a weight based on their relevance to the current token, thus providing salient information about distant tokens.

[0073] According to one example, the ML1 prediction model includes multihead attention layers and feedback attention layers (or "feedforward" attention layers). The attention heads allow the ML1 model to understand the relationships between language elements in a given context, while the feedback attention layers transform these representations. This transformer-based neural network can comprise a stack of multiple transformer layers, each layer constituting an information processing block that progressively improves the understanding of the text being processed.

[0074] According to one example, the transformers of the ML1 model are configured to implement an attention mechanism. Under this mechanism, the attention heads force the transformers to focus on (i.e., give more importance to) specific parts of the input text during the calculation of the output. In this way, the ML1 transcription model is able to analyze and identify long-distance dependencies between language elements (words, for example) and take into account the context necessary to generate a coherent output.

[0075] According to one example, the S6 determination of the DT2 transcription data comprises: - the encoding of the DTI audio data by using a spectrogram image and a convolution operation followed by multiple encoding operations by transformers; and - decoding by a sequence-by-sequence decoding model that predicts a next token (e.g., a word or character) from a transcript and incorporates audio information using an attention mechanism.

[0076] During a sending step S8 ([Fig. 2]), the AGI virtual assistant sends the ML2 massive language model a prompt RQ1 to trigger the execution of function FN1. This prompt RQ1 specifies function FN1 in association with the DT2 transcription data, which serves as the context for performing function FN1. Thus, the DT2 transcription data can define a context in which function FN1 must be performed by the ML2 massive language model.

[0077] In response to the received instruction RQ1, the massive language model ML2 can then perform the function FN1. This function FN1 includes, for example, the generation of a summary (or synthesis, or report, or account) of the speech activity 8 contained in the DTI audio data for a given speaker UR1, for a plurality of speakers UR1, or for all speakers URL. A user, for example One of the UR1 speakers, or a third party, can thus quickly obtain a summary of the discussions taking place between the UR1 speakers during a meeting. This summary can, for example, highlight information relating to the vocal activity of a specific UR1 speaker.

[0078] The FN1 function may, however, vary depending on the case and may, for example, include an automatic translation function or any other appropriate function that can be executed by a massive language model.

[0079] The invention advantageously allows for the identification of speakers UR1 in audio data representative of a speech exchange, without the need for external information about the speakers (no need for an external audio source, audio sample, or predefined personal information). The identity of the respective speaker is included in the transcription data in association with each SGI audio segment transcription, which allows the ML2 massive language model to differentiate the SGI audio segment transcriptions according to their respective speaker. The addition of the speaker identifiers ID1 to the DT2 transcription data enriches the context provided to the ML2 massive language model, enabling it to perform the FN1 function more reliably and efficiently. A massive language model is not capable of distinguishing different speakers without these speaker identifiers IDL.

[0080] Thanks to the invention, tasks normally performed by people, for example employees or others, can be advantageously automated, thus achieving significant savings in time and human resources, while requiring moderate processing resources. For example, the invention makes it possible to automate note-taking during meetings. Minutes can be generated automatically and distributed to interested parties with significant gains in time and efficiency.

[0081] It is advantageously possible to implement the invention without the need to record an audio conversation, for example by using an input audio stream, which makes this system suitable for certain situations where recording a conversation is undesirable. The method of the invention can be carried out in real time from such an audio stream.

[0082] According to a particular example, during the processing procedure described above with reference to [Fig. 2], the AGI virtual assistant obtains a number of speakers UR1 participating in the DTI audio data. The AGI virtual assistant then performs the assignment S4 ([Fig. 2]) based on the number of speakers. In this way, the accuracy of the diarization process, and therefore the performance of the AGI virtual assistant, can be advantageously improved.

[0083] As a person skilled in the art understands, all embodiments and variations The embodiments and variations described above, some of which have been intentionally simplified for ease of explanation, are merely non-exhaustive examples of how this disclosure can be implemented. In particular, those skilled in the art may consider any adaptation or combination of the embodiments and variations described above to meet a specific need.

[0084] The present invention is therefore not limited to the embodiments described above but extends in particular to the previously described methods which would include secondary steps without falling outside the scope of the present invention. The same would apply to the respective devices configured for implementing such methods.

Claims

Claims

1. Processing method implemented by a virtual assistant system (AGI), said method comprising: - detection (S2), from audio data (DTI), of audio segments (SGI) each comprising a vocal activity (8) of a speaker (UR1); - assignment (S4) of a speaker identifier (ID1) to each audio segment respectively; - determination (S6) of transcription data (DT2) respectively associating a transcription (DT3) of the audio segments with the speaker identifiers (ID1); and - sending (S8), to a massive language model (ML2), an instruction (RQ1) to trigger the performance of a function (FN1), said instruction specifying said function in association with the transcription data serving as context for performing said function.

2. Method according to claim 1, in which the transcription data (DT2) specified in the instruction (RQ1) defines a context according to which the function must be performed by the massive language model (ML2).

3. Method according to claim 1 or 2, in which the method comprises, prior to the detection (S2) of the audio segments (SGI), obtaining the audio data (DTI) by extraction from at least one audio file or from at least one audio stream.

4. Method according to any one of claims 1 to 3, wherein a diarization process is used for the detection (S2) of the audio segments and the assignment (S4) of the speaker identifiers (ID1).

5. A method according to any one of claims 1 to 4, wherein determining (S6) the transcription data (DT2) comprises: - transcribing the audio segments (SGI) into text data (DT3) from a voice-to-text transcription model (ML1); and - encoding the text data (DT3) in association with the respective speaker identifiers (ID1) in the transcription data (DT2).

6. Method according to any one of claims 1 to 5, in which the function (FN1) specified in the instruction comprises: - generation of a summary of the voice activity (8) contained in the audio data (DTI) for a given speaker (UR1).

7. A method according to any one of claims 1 to 6, wherein the method comprises obtaining a number of speakers (UR1) involved in the audio data (DTI), in which the assignment (S4) of a speaker identifier (ID1) to the audio segments (SGI) is carried out from the number of speakers.

8. Computer program (PG1) comprising instructions for implementing the method according to any one of the preceding claims, when these instructions are executed by a processor (10).

9. Device (DV1) configured to implement a virtual assistant system (AGI) in cooperation with a massive language model (ML2), said device comprising a memory (16) associated with at least one processor (10) configured to implement the steps of the method according to any one of claims 1 to 7.

10. Processing system (SY1) comprising: - the device (DV1) configured to implement a virtual assistant (AGI) according to claim 9; and - a massive language model (ML2) configured to perform a function (FN1) in response to an instruction (RQ1) sent by the device.

Images