A podcast generation method and related apparatus

Abstract

This application discloses a podcast generation method and related apparatus, relating to the field of artificial intelligence technology. Based on podcast configuration information, a central coordinator agent and N anchor agents are automatically created. By invoking the central coordinator agent, the N anchor agents are scheduled to speak alternately in each dialogue round according to their dynamic speaking weights, simulating the dialogue mode of a real human podcast. The dynamic speaking weights comprehensively consider time decay factors, content relevance factors, and repetition penalty factors, ensuring that anchor agents who have not spoken for a long time, and anchor agents whose potential speaking content has high semantic similarity to the current dialogue context, receive more speaking opportunities. The repetition penalty factor suppresses repetitive speaking behavior, reducing the phenomenon of viewpoint loops and content repetition in multi-round dialogues, achieving precise scheduling of multi-round dialogues, and ensuring the structured, coherent, and interactive nature of the dialogue among the N anchor agents.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a podcast generation method and related apparatus. Background Technology

[0002] With the rapid development of artificial intelligence technology, large language models (LLM) and text-to-speech (TTS) technologies are becoming increasingly mature. AI-assisted or automated podcast generation has become a research hotspot in the field of content creation. Traditional podcast production involves complex processes such as scriptwriting, multi-character voice-over, recording, and post-editing, which is time-consuming, labor-intensive, and costly. The application of AI technology provides an effective way to improve podcast production efficiency and lower the production threshold.

[0003] Currently, some AI podcast generation technologies utilize LLM to generate dialogue text and then use TTS technology to synthesize podcasts. However, these solutions only support basic question-and-answer dialogues and lack effective support for complex multi-role interactions. They struggle to simulate the natural and coherent multi-party interactions between multiple hosts in a real podcast, and are prone to content repetition and loose logic in multi-turn dialogue scenarios. This results in generated podcast content that lacks realism and sufficient information, leading to an overall unsatisfactory effect. Summary of the Invention

[0004] In view of the above problems, this application provides a podcast generation method and related apparatus. The specific solution is as follows:

[0005] The first aspect of this application provides a podcast generation method, including:

[0006] Obtain podcast configuration information, which includes: podcast topic and configuration information for N virtual anchors, wherein the configuration information for the virtual anchors includes anchor role information, and N is a positive integer;

[0007] Based on the podcast configuration information, a central coordinator agent, N anchor agents, and a personalized instruction set corresponding to each anchor agent are created.

[0008] The central coordinator agent is invoked to obtain the dynamic speaking weights of N anchor agents in each dialogue round, and to designate the first target anchor agent to speak based on the dynamic speaking weights of the N anchor agents. The dynamic speaking weights are determined by the base weights and weight influence factors. The weight influence factors include at least one of the following: time decay factor, content relevance factor, and repetition penalty factor. The time decay factor is positively correlated with the duration of the anchor agent's inactivity. The content relevance factor represents the semantic similarity between the anchor agent's potential speaking content and the current dialogue context. The repetition penalty factor is negatively correlated with the repetition index of the anchor agent's dialogue content.

[0009] The first target anchor agent is invoked to output dialogue content in this dialogue round based on the corresponding personalized instruction set and the current dialogue context.

[0010] In one possible implementation, the central coordinator agent and the N anchor agents are connected through a graph-structured framework.

[0011] In one possible implementation, after creating a central coordinator agent, N anchor agents, and a personalized instruction set corresponding to each anchor agent based on the podcast configuration information, the podcast generation method further includes:

[0012] Create a graph-structured framework;

[0013] Each node in the graph-structured framework corresponds to the central coordinator agent and N anchor agents, respectively.

[0014] The state in the graph structured framework is a global data pool shared by all nodes, including the dialogue content output by the N anchor agents;

[0015] The edge between the node corresponding to the central coordinator agent and the node corresponding to the anchor agent represents the transmission path of messages and control flow between the central coordinator agent and the anchor agent.

[0016] The conditional logic in the graph structured framework determines the control flow between nodes based on the dynamic speaking weights of the N anchor agents in each dialogue round.

[0017] In one possible implementation, the podcast generation method further includes:

[0018] Using a sliding window approach, the similarity between the K dialogue items in the current window and the historical dialogue items is calculated.

[0019] The identifier of the second target anchor agent corresponding to the dialogue content with a similarity greater than a similarity threshold is determined, and a dialogue repetition signal carrying the identifier of the second target anchor agent is sent to the central coordinator agent.

[0020] In one possible implementation, after sending a dialogue repetition signal to the central coordinator agent, the podcast generation method further includes:

[0021] The central coordinator agent is invoked to increase the historical dialogue content repetition index of the second target anchor agent based on the dialogue repetition signal.

[0022] In one possible implementation, calculating the similarity between the K dialogue items in the current window and the historical dialogue items includes:

[0023] For each of the K dialogue entries, the dialogue entries are preprocessed to obtain a word set;

[0024] Calculate the TF-IDF value of each word in the word set, and the TF-IDF values ​​of each word in the word set form a dialogue content vector;

[0025] Calculate the cosine similarity between the dialogue content vector and the historical dialogue content vector.

[0026] In one possible implementation, the virtual anchor configuration information also includes a TTS engine identifier and voice timbre, and the podcast generation method further includes:

[0027] The TTS engine corresponding to the first target broadcaster intelligent agent is invoked to synthesize the dialogue content into an audio stream in real time using the timbre corresponding to the first target broadcaster intelligent agent.

[0028] The dialogue content is transmitted to the front end in a streaming manner, and the audio stream is also sent to the front end.

[0029] A second aspect of this application provides a computer program product including computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the podcast generation method of the first aspect or any implementation thereof.

[0030] A third aspect of this application provides an electronic device, comprising at least one processor and a memory connected to the processor, wherein:

[0031] The memory is used to store computer programs;

[0032] The processor is used to execute the computer program to enable the electronic device to implement the podcast generation method of the first aspect or any implementation thereof.

[0033] A fourth aspect of this application provides a computer storage medium carrying one or more computer programs that, when executed by an electronic device, enable the electronic device to perform a podcast generation method as described in the first aspect or any implementation thereof.

[0034] Using the above technical solution, this application provides a podcast generation method and related apparatus that automatically creates a central coordinator agent and N anchor agents based on podcast configuration information. By calling the central coordinator agent, the N anchor agents are scheduled to speak alternately in each dialogue round according to their dynamic speaking weights, simulating the dialogue mode of a real human podcast. The dynamic speaking weights comprehensively consider time decay factors, content relevance factors, and repetition penalty factors to ensure that anchor agents who have not spoken for a long time and anchor agents whose potential speaking content has a high semantic similarity to the current dialogue context get more speaking opportunities. The repetition penalty factor suppresses repetitive speaking behavior, reduces the phenomenon of viewpoint loops and content repetition in multi-round dialogues, and avoids the problems of disordered, repetitive, or off-topic podcast content. This achieves precise scheduling of multi-round dialogues and ensures the structured, coherent, and interactive nature of the dialogue among the N anchor agents. Attached Figure Description

[0035] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0036] Figure 1 A flowchart illustrating a podcast generation method provided in this application embodiment;

[0037] Figure 2 A flowchart illustrating another podcast generation method provided in this application embodiment;

[0038] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0039] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.

[0040] The embodiments of this application will now be described with reference to the accompanying drawings. Those skilled in the art will recognize that, with technological advancements and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are equally applicable to similar technical problems.

[0041] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of elements is not necessarily limited to those elements, but may include other elements not explicitly listed or inherent to those processes, methods, products, or apparatuses.

[0042] This application provides a podcast generation method. The podcast generation method of this application embodiment will be described in detail below with reference to the accompanying drawings.

[0043] Reference Figure 1 , Figure 1 This is a flowchart illustrating a podcast generation method provided in an embodiment of this application, as shown below. Figure 1 As shown in the embodiment of this application, a podcast generation method may include steps 101 to 104, which are described in detail below.

[0044] 101: Get podcast configuration information, which includes: podcast topic and configuration information for N virtual anchors.

[0045] Specifically, it retrieves podcast configuration information entered by the user through the web interface, where N is a positive integer.

[0046] The configuration information for a virtual anchor includes: anchor role information, which includes: name, gender, personality traits, etc.

[0047] 102: Create a central coordinator agent, N anchor agents, and a personalized instruction set for each anchor agent based on the podcast configuration information.

[0048] The Large Language Model (LLM) is initialized based on the podcast configuration information, creating a central coordinator agent (Supervisor Agent), N host agents, and a personalized instruction set (Prompt) for each host agent. The central coordinator agent does not directly generate dialogue content, but acts as a "director" and "host," responsible for the macro-level scheduling and management of the host agents.

[0049] The personalized instruction sets corresponding to different anchor agents are not exactly the same. The personalized instruction set of the anchor agent not only defines the role of the anchor agent, but also describes the personality traits and speaking style of the anchor agent. The personalized instruction set can also be dynamically adjusted.

[0050] For example, the common part of the personalized instruction sets corresponding to N anchor agents may include:

[0051] It uses natural, conversational expressions, making it suitable for TTS (Text-to-Speech) playback.

[0052] Use appropriate interjections (such as "um," "ah," "ne") to increase realism;

[0053] Provide smooth topic transition words (such as "Speaking of which", "By the way");

[0054] Expand and explain the knowledge based on the discussion content;

[0055] Maintain a clear flow in the dialogue, including an opening, in-depth discussion, and a summary.

[0056] Respond to other broadcasters' opinions and promote interaction;

[0057] Control the length of your speech and avoid excessively long monologues;

[0058] Each speech should present a new viewpoint to avoid repetition;

[0059] Output the dialogue content directly without adding any extra information.

[0060] This meticulous Prompt engineering allows each anchor agent to simulate a distinctive speaking style, greatly enhancing the naturalness and appeal of the generated podcasts.

[0061] 103: Invoke the central coordinator agent to obtain the dynamic speaking weights of N anchor agents in each dialogue round, and specify the first target anchor agent to speak based on the dynamic speaking weights of the N anchor agents.

[0062] The central coordinator agent designates the first speaker agent to initiate the first round of dialogue. The central coordinator agent can designate the first speaker agent in several ways, such as randomly, based on podcast configuration information, or based on the base weights of N speaker agents.

[0063] The selected anchor agent generates a dialogue based on its personalized instruction set and the current dialogue context.

[0064] The central coordinator agent acquires the dynamic speaking weights of N anchor agents in each dialogue round, and designates the first target anchor agent to speak based on the dynamic speaking weights of the N anchor agents. The dynamic speaking weights are determined by the base weights and weight influence factors. The weight influence factors include at least one of the following: time decay factor, content relevance factor, and repetition penalty factor. The time decay factor is positively correlated with the anchor agent's inactive time. The content relevance factor represents the semantic similarity between the anchor agent's potential speaking content and the current dialogue context. The repetition penalty factor is negatively correlated with the anchor agent's historical dialogue content repetition index.

[0065] For example, suppose there is Anchor AI Agent In each round of dialogue intelligent agent Dynamic speech weight It can be represented as:

[0066] ;

[0067] in, The base weight represents the initial base speaking weight that each anchor agent possesses.

[0068] This is the time decay factor, representing the time since the last speech. Since then, the anchor intelligent agent The willingness to speak has accumulated. With As the duration increases, the time decay factor increases, encouraging anchor agents who have not spoken for a long time to speak. For example, , It is a positive decay constant.

[0069] As a content relevance factor, evaluate the anchor's intelligent agent. With the current dialogue context The relevance of the statements made in China. It can be calculated The semantic similarity between potential speech content and the current dialogue context is obtained. Semantic similarity can be calculated based on word embeddings or topic models.

[0070] As a repetition penalty factor, when the anchor agent... Historical statement When duplicate content or potential statements are flagged by the duplicate detection module, the weight of the duplicate penalty factor is reduced. For example, , It is a live streamer intelligent agent The historical dialogue content duplication index, based on... Potential content of speech and The cosine similarity of the historical speech content is calculated, or based on... The content of the speech and The cosine similarity of the historical speech content was calculated. It is a positive penalty constant.

[0071] When selecting the next speaking agent, the central coordinator agent does so based on the dynamic weights of the N main speaking agents. Perform a probabilistic selection. Select the anchor agent. probability It can be calculated using the Softmax function:

[0072] ;

[0073] Choose probability The highest-ranking anchor agent is designated as the primary target anchor agent. Through this dynamic weight adjustment and probability selection mechanism, the balance of the dialogue can be ensured, a single anchor agent can be prevented from dominating the screen, and anchor agents can be encouraged to make valuable remarks at appropriate times.

[0074] 104: Invoke the first target anchor AI agent to output dialogue content in this dialogue round based on the corresponding personalized instruction set and the current dialogue context.

[0075] Each broadcaster agent corresponds to a large language model. The personalized instruction set corresponding to the first target broadcaster agent and the current dialogue context are input into the large language model of the first target broadcaster agent, so that the large language model outputs the dialogue content.

[0076] This embodiment provides a podcast generation method that automatically creates a central coordinator agent and N host agents based on podcast configuration information. By calling the central coordinator agent, the N host agents are scheduled to speak alternately in each dialogue round according to their dynamic speaking weights, simulating the dialogue mode of a real human podcast. The dynamic speaking weights comprehensively consider time decay factors, content relevance factors, and repetition penalty factors to ensure that host agents who have not spoken for a long time and host agents whose potential speaking content has high semantic similarity to the current dialogue context get more speaking opportunities. The repetition penalty factor suppresses repetitive speaking behavior, reduces the phenomenon of viewpoint loops and content repetition in multi-round dialogues, and avoids the problems of disordered, repetitive, or off-topic podcast content. This achieves precise scheduling of multi-round dialogues and ensures the structured, coherent, and interactive nature of the dialogue among the N host agents.

[0077] In one possible implementation, the central coordinator agent and the N anchor agents are connected through a graph structure framework (such as LangGraph). The scheduling logic of the central coordinator agent is implemented through the functions of the graph structure framework, integrating the central coordinator agent and the N anchor agents into a compilable graph structure and supporting flexible scheduling strategies and message passing.

[0078] In the above embodiments, after step 102, a graph-structured framework is created to abstract the dialogue flow between N anchor agents into a directed graph, wherein:

[0079] In the graph-structured framework, each node corresponds to the central coordinator agent and N anchor agents, respectively.

[0080] In the graph-structured framework, the State is a global data pool shared by all nodes, including the dialogue content output by N anchor agents, that is, all historical dialogue content output by N anchor agents, enabling the central coordinator agent to make decisions based on the complete context and ensuring the coherence of the dialogue.

[0081] The edges between the nodes corresponding to the central coordinator agent and the nodes corresponding to the anchor agent represent the transmission paths of messages and control flows between the central coordinator agent and the anchor agent. For example, the output of the central coordinator agent can be used as the input of the next anchor agent.

[0082] Conditional logic in a graph-structured framework supports defining loops and conditional branches, which is crucial for implementing complex dialogue logic. For example, conditional logic determines the control flow between nodes based on the dynamic speaking weights of N host agents in each dialogue round. When dialogue repetition is detected, the central coordinator agent can trigger a loop to re-evaluate and designate the next host agent to speak.

[0083] In one possible implementation, the graph-structured framework also includes duplicate detection of dialogue content corresponding to nodes.

[0084] To ensure the quality and information content of the generated dialogue, this application embodiment incorporates a dialogue duplication detection mechanism, which effectively avoids the problem of content repetition or topic looping in long-term dialogues by virtual anchors.

[0085] In one possible implementation, a sliding window approach is used to acquire K dialogue pieces within the current window. K is set according to the specific application scenario; for example, K could be set to 3. Then, the similarity between the K dialogue pieces in the current window and historical dialogue pieces is calculated. The identifier of the second target anchor agent corresponding to dialogue pieces with a similarity greater than a similarity threshold (e.g., 0.6) is determined, and a dialogue repetition signal carrying the identifier of the second target anchor agent is sent to the central coordinator agent.

[0086] Furthermore, the central coordinator agent can be invoked to increase the historical dialogue content repetition index of the second target anchor agent based on the dialogue repetition signal.

[0087] It can also invoke the central coordinator agent to prompt the currently speaking anchor agent to re-output the dialogue content based on the dialogue repetition signal, or to reassign another anchor agent to speak, thereby avoiding the output of duplicate content and improving the quality of the dialogue.

[0088] There are several ways to calculate the similarity between the K dialogues in the current window and the historical dialogues. Two methods are illustrated below.

[0089] Method 1:

[0090] The Jaccard similarity index is used to quantify the similarity between dialogue texts.

[0091] Jaccard similarity (Jaccard Index) is a metric for measuring the similarity between two sets, defined as the size of their intersection divided by the size of their union. In this embodiment, Jaccard similarity is used to evaluate the similarity between dialogue content and historical dialogue content. The formula for calculating Jaccard similarity is:

[0092] ;

[0093] in, and Representing text respectively and text The word set obtained after preprocessing (such as word segmentation, stop word removal, etc.). In this implementation... This represents the set of words corresponding to a single dialogue item within the current window. This refers to the set of words corresponding to the content of historical dialogues.

[0094] express and The size of the intersection (i.e. the number of common words). express and The size of the union of the two sets (i.e., the total number of all unique words). The Jaccard similarity score ranges from [value missing]. The higher the value, the higher the similarity, indicating that the two texts have more common words and are more similar in content.

[0095] Method 2:

[0096] To more accurately assess the similarity between dialogue texts, this embodiment combines TF-IDF (Term Frequency-Inverse Document Frequency) weights and cosine similarity to calculate text similarity, specifically including the following steps A1-A3:

[0097] A1: For each of the K dialogues, preprocess the dialogue content to obtain a word set;

[0098] Preprocessing of dialogue content includes word segmentation and removal of stop words.

[0099] A2: Calculate the TF-IDF value of each word in the word set. The TF-IDF values ​​of each word in the word set form the dialogue content vector.

[0100] The TF-IDF value is used to assess the importance of a word in a set of words to a set of K words. A word's importance increases directly with its frequency in the set of words, but decreases inversely with its frequency in the K word sets. The formula for calculating TF-IDF is as follows:

[0101] Term Frequency (TF):

[0102] ;

[0103] Inverse document frequency (IDF):

[0104] ;

[0105] in, Let K be the value.

[0106] TF-IDF value:

[0107] .

[0108] A vector is formed by the TF-IDF values ​​of each word in a dialogue within the current window. The vector is composed of the TF-IDF values ​​of each word in the historical dialogue content. .

[0109] A3: Calculate the cosine similarity between the dialogue content vector and the historical dialogue content vector.

[0110] vector sum vector The formula for calculating the cosine similarity between them is as follows:

[0111] ;

[0112] in, and They are vectors and The Each component (corresponding to the TF-IDF value of a given word). The range of cosine similarity values ​​is... The closer the value is to 1, the higher the similarity; the closer it is to -1, the lower the similarity; and the closer it is to 0, the more independent they are.

[0113] The above two methods are merely examples, and this application is not limited to them. In one possible implementation, the final similarity can also be determined by combining Jaccard similarity and TF-IDF cosine similarity, so as to more comprehensively and accurately assess the repetition of dialogue content and thus improve the robustness of repetition detection.

[0114] In one possible implementation, the virtual anchor configuration information also includes a TTS engine identifier and voice timbre. For example, a deep male voice can be configured for a male virtual anchor, and a clear female voice can be configured for a female virtual anchor. It is even possible to select a more expressive voice timbre based on the personality characteristics of the virtual anchor. That is, different anchor agents use different TTS engines and voice timbres for speech synthesis. This embodiment can integrate and switch between multiple TTS engines and allow users to select different voice timbres for each AI anchor to meet diverse auditory experience needs.

[0115] Please see Figure 2 The flowchart shown in this application illustrates a podcast generation method that includes the following steps 201-206:

[0116] 201: Retrieve podcast configuration information, which includes: podcast topic and configuration information for N virtual anchors;

[0117] 202: Create a central coordinator agent, N anchor agents, and a personalized instruction set for each anchor agent based on the podcast configuration information;

[0118] 203: Invoke the central coordinator agent to obtain the dynamic speaking weights of N anchor agents in each dialogue round, and specify the first target anchor agent to speak based on the dynamic speaking weights of the N anchor agents;

[0119] 204: Invoke the first target anchor AI agent to output dialogue content in this dialogue round based on the corresponding personalized instruction set and the current dialogue context;

[0120] 205: Call the TTS engine corresponding to the first target anchor agent, and use the voice corresponding to the first target anchor agent to perform real-time streaming speech synthesis of the dialogue content to generate an audio stream;

[0121] 206: Stream the conversation to the front end, and simultaneously send the audio stream to the front end.

[0122] This embodiment provides a podcast generation method whereby, after the first target anchor agent outputs dialogue content, it simultaneously performs real-time streaming speech synthesis through a corresponding TTS engine, and transmits the text and audio streams to the front end in a streaming manner, achieving an end-to-end real-time experience of "generating, synthesizing, and playing simultaneously." Users do not need to wait for all content to be generated before listening, thus improving the user experience.

[0123] In one possible implementation, a central coordinating agent could be invoked to monitor the dialogue content, ensuring it stays on the pre-defined topic. When the dialogue deviates, stalls, or reaches a stalemate, the Supervisor immediately intervenes, using topic redirection strategies to guide the conversation back to the core issue or towards a new angle of discussion. Furthermore, upon receiving a warning from the dialogue quality control module (such as duplicate detection), the Supervisor activates a crisis intervention mechanism, such as pausing the current speaker, sending correction instructions to the relevant host, or directly switching speakers to quickly correct any deviations in the dialogue.

[0124] The central coordinator agent controls the overall length and pace of the conversation by adjusting the frequency and number of speaking turns. It adaptively adjusts the pace based on the user-defined total podcast duration, the current progress of the conversation, and the richness of the content. For example, it speeds up the pace when there is abundant content but limited time, and slows down the pace when there is less content but more time, encouraging the host to explore or expand on the topic. This adaptive mechanism makes the podcast's rhythm more natural and in line with listeners' expectations.

[0125] This application embodiment can have multiple application scenarios. In addition to general podcast generation, this application embodiment can also be extended to:

[0126] News Broadcast: Real-time generation of multi-anchor news broadcasts provides a more vivid auditory experience.

[0127] Online education: Simulates multiple teachers to conduct course explanations or Q&A sessions, enhancing learning interactivity.

[0128] Audiobook creation: Generating audio content with multi-character dialogues for novels, scripts, etc.

[0129] Virtual Customer Service / Assistant: Build more human-like and interactive multi-role AI customer service or virtual assistants.

[0130] This application also provides an electronic device in its embodiments. (See reference...) Figure 3 As shown, it illustrates a structural schematic diagram suitable for implementing the electronic device in the embodiments of this application. Figure 3 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0131] like Figure 3 As shown, the electronic device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 301, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 302 or a program loaded from a storage device 308 into a random access memory (RAM) 303. When the electronic device is powered on, the RAM 303 also stores various programs and data required for the operation of the electronic device. The processing unit 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0132] Typically, the following devices can be connected to I / O interface 305: input devices 306 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 307 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 308 including, for example, memory cards, hard drives, etc.; and communication devices 309. Communication device 309 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 3 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or have alternatively.

[0133] This application also provides a computer program product including computer-readable instructions, which, when executed on an electronic device, cause the electronic device to implement any of the podcast generation methods provided in this application.

[0134] This application also provides a computer-readable storage medium that carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any of the podcast generation methods provided in this application.

[0135] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the device embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0136] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0137] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0138] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

Claims

1. A podcast generation method, characterized in that, include: Obtain podcast configuration information, which includes: podcast topic and configuration information for N virtual anchors, wherein the configuration information for the virtual anchors includes anchor role information, and N is a positive integer; Based on the podcast configuration information, a central coordinator agent, N anchor agents, and a personalized instruction set corresponding to each anchor agent are created. The central coordinator agent is invoked to obtain the dynamic speaking weights of N anchor agents in each dialogue round, and to designate the first target anchor agent to speak based on the dynamic speaking weights of the N anchor agents. The dynamic speaking weights are determined by the base weights and weight influence factors. The weight influence factors include: time decay factor, content relevance factor, and repetition penalty factor. The time decay factor is positively correlated with the anchor agent's inactive time, representing the accumulated speaking intention of the anchor agent since the last speaking time. The content relevance factor represents the semantic similarity between the anchor agent's potential speaking content and the current dialogue context. The repetition penalty factor is negatively correlated with the repetition index of the anchor agent's dialogue content. The first target anchor AI agent is invoked to output dialogue content in this dialogue round based on the corresponding personalized instruction set and the current dialogue context.

2. The podcast generation method according to claim 1, characterized in that, The central coordinator agent and the N anchor agents are connected through a graph-structured framework.

3. The podcast generation method according to claim 2, characterized in that, After creating a central coordinator agent, N anchor agents, and a personalized instruction set for each anchor agent based on the podcast configuration information, the podcast generation method further includes: Create a graph-structured framework; Each node in the graph-structured framework corresponds to the central coordinator agent and N anchor agents, respectively. The state in the graph structured framework is a global data pool shared by all nodes, including the dialogue content output by the N anchor agents; The edge between the node corresponding to the central coordinator agent and the node corresponding to the anchor agent represents the transmission path of messages and control flow between the central coordinator agent and the anchor agent. The conditional logic in the graph structured framework determines the control flow between nodes based on the dynamic speaking weights of the N anchor agents in each dialogue round.

4. The podcast generation method according to claim 1, characterized in that, The podcast generation method also includes: Using a sliding window approach, the similarity between the K dialogue items in the current window and the historical dialogue items is calculated respectively; The identifier of the second target anchor agent corresponding to the dialogue content with a similarity greater than a similarity threshold is determined, and a dialogue repetition signal carrying the identifier of the second target anchor agent is sent to the central coordinator agent.

5. The podcast generation method according to claim 4, characterized in that, After sending a dialogue repetition signal to the central coordinator agent, the podcast generation method further includes: The central coordinator agent is invoked to increase the historical dialogue content repetition index of the second target anchor agent based on the dialogue repetition signal.

6. The podcast generation method according to claim 4, characterized in that, The calculation of the similarity between the K dialogue entries in the current window and the historical dialogue entries includes: For each of the K dialogue entries, the dialogue entries are preprocessed to obtain a word set; Calculate the TF-IDF value of each word in the word set, and the TF-IDF values ​​of each word in the word set form a dialogue content vector; Calculate the cosine similarity between the dialogue content vector and the historical dialogue content vector.

7. The podcast generation method according to claim 1, characterized in that, The virtual anchor configuration information also includes a TTS engine identifier and voice timbre, and the podcast generation method further includes: The TTS engine corresponding to the first target broadcaster intelligent agent is invoked to synthesize the dialogue content into an audio stream in real time using the timbre corresponding to the first target broadcaster intelligent agent. The dialogue content is transmitted to the front end in a streaming manner, and the audio stream is also sent to the front end.

8. A computer program product, characterized in that, Includes computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the podcast generation method as described in any one of claims 1 to 7.

9. An electronic device, characterized in that, It includes at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is used to execute the computer program to enable the electronic device to implement the podcast generation method as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that, The storage medium carries one or more computer programs that, when executed by an electronic device, enable the electronic device to implement the podcast generation method as described in any one of claims 1 to 7.

Images