An Automatic EQ Tuning System and Method Based on a Large Language Model

Abstract

The application discloses an automatic EQ tuning system and method based on a large language model, and the system comprises a user input module, an information collection module, a ReAct inference engine, a tool calling module and an output module; the user input module receives a tuning requirement input by a user and acquires text data; the information collection module receives the text data and acquires relevant information of the corresponding user; the ReAct inference engine is connected with the information collection module, processes the received tuning requirement and the relevant information of the corresponding user based on a large language model, and outputs a final EQ parameter set; the tool calling module is coupled with the ReAct inference engine, calls a third-party tool to support information processing of the ReAct inference engine; and the output module processes and outputs the EQ parameter set. The application realizes efficient and accurate tuning, and reduces the dependence on professional acoustic knowledge.

Description

Technical Field

[0001] This invention relates to the field of audio processing technology, and in particular to an automatic EQ tuning system and method based on a large language model. Background Technology

[0002] With the rapid development of audio processing technology, sound quality optimization is becoming increasingly important in music production, film and television post-production, live streaming, and consumer electronics. Equalizers (EQ), as core tools in audio processing, optimize audio signals by adjusting the gain of different frequency bands to meet specific scenarios or user needs.

[0003] However, traditional EQ tuning mainly relies on manual adjustments by professional audio engineers, which requires a high level of expertise. Operators need to have extensive audio knowledge and experience, making it difficult for ordinary users to achieve high-quality sound optimization, resulting in a complex and inefficient tuning process.

[0004] In addition, existing EQ tuning solutions often fail to fully consider the acoustic characteristics of the headphones themselves, such as frequency response curves, impedance differences, or sound field characteristics, resulting in a discrepancy between the tuning results and the actual listening experience, making it difficult to adapt to the personalized needs of different headphone models. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides an automatic EQ tuning system and method based on a large language model. By leveraging the semantic understanding capabilities of the large language model and combining it with the acoustic characteristics of headphones and audio signal processing technology, it achieves efficient and high-quality sound optimization while reducing reliance on specialized knowledge.

[0006] This invention provides an automatic EQ tuning system based on a large language model, the specific technical solution of which is as follows:

[0007] It includes a user input module, an information collection module, a ReAct inference engine, a tool invocation module, and an output module;

[0008] The user input module receives the user's tuning request and obtains the text data;

[0009] The information acquisition module receives the text data and obtains relevant information of the corresponding user; the ReAct inference engine is connected to the information acquisition module, processes the received tuning requirements and relevant information of the corresponding customer based on the large language model, and outputs the final EQ parameter set.

[0010] The tool invocation module is coupled with the ReAct inference engine and invokes third-party tools to support the information processing of the ReAct inference engine.

[0011] The output module processes and outputs the EQ parameter set.

[0012] Furthermore, the tools that the tool calling module can call include frequency response curve acquisition tools, song style query tools, and EQ tuning knowledge base tools.

[0013] Furthermore, the frequency response curve acquisition tool includes a first vector database, a retrieval and download module, and a uniform sampling unit;

[0014] The first vector database stores mapping data between headphone models and their measured frequency response curve file addresses;

[0015] The retrieval and download module is connected to the first vector database, retrieves the corresponding measured frequency response curve file address according to the headphone model, and downloads the measured frequency response curve file.

[0016] The uniform sampling unit is connected to the retrieval and download module, receives the measured frequency response curve file, and downsamples the original data points of the measured frequency response curve to a fixed number.

[0017] Furthermore, the EQ tuning knowledge base tool includes a second vector database, an embedding model, and a reordering model;

[0018] The second vector database stores mapping data between song styles and tuning suggestions;

[0019] The embedding model retrieves tuning suggestions with high similarity.

[0020] The re-ranking model scores the semantic relevance of the vector retrieval results and outputs several tuning suggestions from high to low relevance scores.

[0021] Furthermore, the output module includes a role-playing module, which encapsulates the final EQ parameter set into a natural language response with a preset persona.

[0022] Furthermore, the output module also includes a front-end display module, which parses the final EQ parameter set and draws an interactive EQ curve in real time.

[0023] Furthermore, the system also includes a DSP audio processing module. The ReAct inference engine is connected to the front-end display module and transmits the final EQ parameter set to the front-end display module for display output. After receiving the interactive command, the front-end display module sends the final EQ parameter set to the digital signal processor to update the audio filter.

[0024] This invention also discloses an automatic EQ tuning method based on a large language model. The automatic EQ tuning method based on a large language model, as described above, is characterized by comprising:

[0025] S1: Receive user input of tuning requirements;

[0026] S2: Obtain relevant information about the current user, including the model of the headphones currently connected to the client, the information of the song currently being played, the current EQ, and historical conversation messages;

[0027] S3: Perform multiple rounds of iterative reasoning through a large language model until the number of iterations is met or the final result label is triggered, and generate the final EQ parameter set;

[0028] S4: Send the final EQ parameter set to the digital signal processor to update the audio filter.

[0029] Furthermore, in step S3, the large language model adopts an LLM model, generating the name and parameters of the tool to be called in each round of inference, calling the corresponding tool to obtain the return result, and using the result as the input for the next round.

[0030] Furthermore, after step S3, the method further includes: encapsulating the final EQ parameter set into a natural language response with a preset persona for output, and drawing an EQ curve based on the final EQ parameter set for output.

[0031] The beneficial effects of this invention are as follows:

[0032] 1. This invention supports natural language input through a user input module. Combined with the semantic understanding capabilities of LLM, it enables tuning to be driven by natural language. Users can achieve high-quality sound optimization without professional audio knowledge, eliminating the dependence on professional acoustic knowledge. At the same time, combined with the ReAct inference engine, it integrates user needs, headphone characteristics, song style and historical dialogue through an iterative cycle of thinking-action-observation to achieve dynamic adjustment of EQ parameters and generate accurate EQ parameters.

[0033] 2. This invention uses a tool to call a module to obtain the frequency response curve data of the headphones, and uses LLM to dynamically adjust the EQ parameters in combination with the characteristics of the headphones to ensure that the tuning result is highly matched with the acoustic characteristics of the headphones, thereby improving the personalized effect of sound quality. It also uses a tool to call the song style information and generate style-related EQ suggestions to ensure that the tuning matches the music content, enhance the bass power and vocal clarity, and improve the targeted listening experience.

[0034] 3. This invention presents the adjusted EQ curve in real time through the front-end display module, enabling users to intuitively understand and fine-tune the parameters. It also achieves efficient sound quality optimization on resource-constrained devices such as mobile phones through real-time processing of the audio processing module. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0036] The technical solutions in the embodiments of the present invention are clearly and completely described in the following description. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0037] In the description of the embodiments of the present invention, it should be noted that the indicated orientation or positional relationship is based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the product of the invention is conventionally placed during use, or the orientation or positional relationship in which those skilled in the art conventionally understand it during use. This is only for the convenience of describing the present invention and simplifying the description, and is not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of the present invention. Furthermore, the terms "first" and "second" are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0038] In the description of the embodiments of the present invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set" and "connection" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0039] Example 1

[0040] Embodiment 1 of the present invention discloses an automatic EQ tuning system based on a large language model, including a user input module, an information acquisition module, a ReAct inference engine, a tool calling module, an output module, and a DSP audio processing module;

[0041] The user input module receives the user's tuning request in natural language form and obtains text data. The user input module captures the user's intent as the basis for all subsequent steps.

[0042] The information collection module receives the text data and obtains relevant information about the corresponding user;

[0043] Specifically, when the information collection module receives text data, it obtains a unique identifier for the user's identity and retrieves relevant information about the user based on the unique identifier.

[0044] The relevant information includes the connected headphone model, current EQ, currently playing song information, and historical conversation messages, etc. The information collection module also includes obtaining a list of tools available in the current system.

[0045] By collecting relevant information, we can enrich the input and ensure that the system does not just deal with isolated problems.

[0046] The ReAct inference engine is connected to the information acquisition module. Based on a large language model, it processes the received tuning requirements and relevant customer information in a multi-round iterative manner of thinking-action-observation, and outputs the final EQ parameter set.

[0047] Specifically, the ReAct inference engine has a maximum number of iterations during iterative processing, and when the maximum number of iterations is reached, it will forcibly enter the role-playing module.

[0048] The ReAct inference engine uses an LLM model. Based on the input: user question + collected relevant information (current Bluetooth headset model used by the user + currently playing song information + current EQ + list of available tools + user's historical conversation messages), the response output is in the following format:

[0049] Thought: xxxxxxx;

[0050] Action: xxxxxxx;

[0051] Action Input: xxxxxxx;

[0052] Final answer: xxxxxxx;

[0053] Direct_answer:xxxxxx;

[0054] Here, Thought represents the thought label, xxxxxxx represents the specific content of the thought analysis; Action represents the tool name, Action Input represents the parameters of the tool, Final_answer represents the final answer obtained after multiple rounds of reasoning, and Direct_answer represents a direct reply without calling the tool.

[0055] The tool invocation module is coupled with the ReAct inference engine, and invokes third-party tools to support the information processing of the ReAct inference engine. When a tool invocation tag is detected, the corresponding tool is invoked and the operation is performed according to the parameters.

[0056] The tools that can be invoked by the tool invocation module include: frequency response curve acquisition tool, song style query tool, and EQ tuning knowledge base tool;

[0057] The frequency response curve acquisition tool is used to query and return the measured frequency response data of the headphone in the range of 20Hz–20kHz based on the headphone model.

[0058] In a preferred embodiment, the frequency response curve acquisition tool includes a first vector database, a retrieval and download module, and a uniform sampling unit;

[0059] The first vector database (weaviate vector database) stores mapping data between headphone models and their measured frequency response curve file addresses;

[0060] Specifically, fields such as headphone model, alias, and the address of frequency response curve file are set in the schema of the vector database to support vector similarity search.

[0061] The retrieval and download module is connected to the first vector database, retrieves the corresponding measured frequency response curve file address according to the headphone model, and downloads the measured frequency response curve file.

[0062] Specifically, when the input includes the headphone model collected from the context, the model is matched by vector retrieval, and the URL of the frequency response curve is returned; the file (usually CSV or TXT, containing frequency-amplitude data points) can be downloaded via HTTP request from the obtained URL, ensuring file integrity verification.

[0063] The uniform sampling unit is connected to the retrieval and download module, receives the measured frequency response curve file, and downsamples the original data points of the measured frequency response curve to a fixed number N, where 20≤N≤30, in order to reduce the token length input to the large language model.

[0064] In this embodiment, the uniform sampling unit selects the N data points at equal intervals on a logarithmic frequency scale;

[0065] Specifically, select 20 representative data points from the downloaded file (e.g., evenly distributed on a logarithmic frequency scale, such as 20Hz, 40Hz, ..., 20kHz); assuming there are 500 data points, take one data point every 25 points, which will accumulate to 20 data points.

[0066] The final result is an array of 20 sampling points (e.g., [(20Hz, -2dB), (40Hz, 0dB), ..., (20kHz, -1dB)]).

[0067] A song style query tool is used to obtain the style tags of the currently playing song;

[0068] Specifically, when parsing the content of the LLM response as follows:

[0069] Action: web_search;

[0070] Action Input: query:{specific song name and style};

[0071] It will call the web_search tool, passing the parameter query:{specific song name and style} to the tool, and then send a search query to get the query results.

[0072] An EQ tuning knowledge base tool is used for vector retrieval of EQ tuning suggestions that match the style tags and natural language tuning requirements.

[0073] The EQ tuning knowledge base tool includes a second vector database, an embedding model, and a reordering model;

[0074] The second vector database stores mapping data between song styles and tuning suggestions;

[0075] The embedding model retrieves tuning suggestions with high similarity. Specifically, when the content of the parsed LLM response is as follows:

[0076] Action: search_eq_tips;

[0077] Action Input: query:{Specific song style EQ tuning suggestions};

[0078] It calls the search_eq_tips tool, which retrieves the query value from the vector database to obtain tuning suggestions with high similarity, and then builds a candidate list.

[0079] The re-ranking model scores the semantic relevance of the vector retrieval results and outputs several tuning suggestions from high to low relevance scores.

[0080] The output module processes and outputs the EQ parameter set;

[0081] Specifically, output is executed when the final answer or a direct reply tag is detected.

[0082] In a preferred embodiment, the output module includes a role-playing module, which encapsulates the final EQ parameter set into a natural language response with a preset persona.

[0083] In a preferred embodiment, the output module further includes a front-end display module. The ReAct inference engine is connected to the front-end display module and is used to parse the final EQ parameter set and draw interactive EQ curves and display EQ buttons in real time.

[0084] After receiving the interactive command (i.e., the user presses the EQ button), the front-end display module sends the final EQ parameter set to the digital signal processor to update the audio filter, that is, adjusts the filter parameters according to the EQ settings (e.g., sets the cutoff frequency or Q value) to ensure real-time or offline audio output optimization.

[0085] Each EQ node in the final EQ parameter set includes at least frequency, Q value, and gain, with the absolute value of single-band gain ≤3dB and the overall minimum gain ≥-12dB.

[0086] In this embodiment, the DSP audio processing module is configured to update the audio filter in real time on the mobile device with a delay of <50ms.

[0087] Example 2

[0088] Embodiment 2 of the present invention discloses an automatic EQ tuning method based on a large language model, which is based on the automatic EQ tuning method based on a large language model described in Embodiment 1, such as... Figure 1 As shown, the specific process is as follows:

[0089] S1: Receives the user's input in natural language form regarding tone requirements;

[0090] S2: Obtain relevant information about the current user, including the model of the headphones currently connected to the client, the information of the song currently being played, the current EQ, and historical conversation messages;

[0091] S3: Perform multiple rounds of iterative reasoning through a large language model until the number of iterations is met or the final result label is triggered, and generate the final EQ parameter set;

[0092] The large language model adopts an LLM model. In each round of inference, the name and parameters of the tool to be called are generated, the corresponding tool is called to obtain the return result, and the result is used as the input for the next round.

[0093] Each EQ node in the set contains at least a frequency, a Q value, and a gain, and the absolute value of the gain of a single frequency band is ≤3dB, and the overall minimum gain is ≥-12dB;

[0094] The tools include at least:

[0095] Frequency response curve acquisition tool, used to return the measured frequency response data of headphones in the range of 20Hz–20kHz;

[0096] A song style lookup tool used to return the style tags of the currently playing song;

[0097] An EQ tuning knowledge base tool for vector-based EQ tuning suggestions that match the style tags and natural language tuning requirements.

[0098] In a preferred embodiment, the final EQ parameter set is encapsulated into a natural language response with a preset persona and output, and an EQ curve is plotted and output based on the final EQ parameter set.

[0099] S4: Send the final EQ parameter set to the digital signal processor to update the audio filter.

[0100] This invention is not limited to the specific embodiments described above. The invention extends to any new feature or combination disclosed in this specification, as well as any new method or process step or combination disclosed herein.

Claims

1. A large language model-based automatic EQ tuning system, characterized in that, The system comprises a user input module, an information collection module, a ReAct inference engine, a tool calling module, and an output module. The user input module receives the user's input tuning requirements and obtains text data. The information collection module receives the text data and obtains relevant information of the corresponding user. The ReAct inference engine is connected with the information collection module, and based on a large language model, processes the received tuning requirements and relevant information of the corresponding user, and outputs a final EQ parameter set. The tool calling module is coupled with the ReAct inference engine, and calls third-party tools to support information processing of the ReAct inference engine; the tools that can be called by the tool calling module include a frequency response curve acquisition tool, a song style query tool, and an EQ tuning knowledge base tool. The frequency response curve acquisition tool comprises a first vector database, a retrieval and download module, and a uniform sampling unit; the first vector database stores mapping data of earphone models and their measured frequency response curve file addresses; the retrieval and download module is connected with the first vector database, retrieves the corresponding measured frequency response curve file address according to the earphone model, and downloads and obtains the measured frequency response curve file; the uniform sampling unit is connected with the retrieval and download module, receives the measured frequency response curve file, and down-samples the original data points of the measured frequency response curve to a fixed number. The EQ tuning knowledge base tool comprises a second vector database, an embedding model, and a reordering model; the second vector database stores mapping data of song styles and tuning suggestions; the embedding model retrieves and obtains tuning suggestions with high similarity; the reordering model scores the vector retrieval results in terms of semantic relevance and outputs a number of tuning suggestions from high to low in terms of relevance scores. The output module processes and outputs the EQ parameter set.

2. The automatic EQ tuning system based on large language model according to claim 1, wherein, The output module comprises a role-playing module, which encapsulates the final EQ parameter set into a natural language reply with a preset persona.

3. The automatic EQ tuning system based on large language model according to claim 1, wherein, The output module further comprises a front-end display module, which analyzes the final EQ parameter set and real-time draws an interactive EQ curve.

4. The automatic EQ tuning system based on large language model according to claim 3, characterized in that, The system further comprises a DSP audio processing module, the ReAct inference engine is connected with the front-end display module, transmits the final EQ parameter set to the front-end display module for display and output, and after receiving an interactive instruction, the front-end display module downloads the final EQ parameter set to the DSP audio processing module to update the audio filter.

5. An automatic EQ tuning method based on a large language model, characterized in that, The automatic EQ tuning system based on a large language model according to any one of claims 1-4, comprising: S1: receiving user input tuning requirements; S2: obtaining relevant information of the current user, including the earphone model connected by the current user, the song information currently played, the current EQ, and the historical dialogue messages; S3: performing multi-round iterative reasoning through a large language model until the number of iterations is met or the final result label is triggered, to generate a final EQ parameter set; S4: downloading the final EQ parameter set to a digital signal processor to update the audio filter.

6. The automatic EQ tuning method based on a large language model according to claim 5, characterized in that, In step S3, the large language model adopts an LLM model, generates a tool name and parameters to be called in each round of reasoning, calls the corresponding tool to obtain a return result, and takes the result as a new round of input.

7. The automatic EQ tuning method based on a large language model according to claim 6, characterized in that, After step S3, it further includes: encapsulating the final set of EQ parameters into a natural language reply with a preset human configuration for output, and drawing an EQ curve based on the final set of EQ parameters for output.

Images