Method of generating impulse responses
Patent Information
- Application Number
- TW114106916
- Authority / Receiving Office
- TW · TW
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2025-02-25
- Publication Date
- 2026-09-01
- Estimated Expiration
- 2045-02-24
AI Technical Summary
Current methods for generating speaker impulse responses rely on traditional measurement and recording, making it difficult to achieve rapid and accurate simulation for personalized needs.
A method using deep learning models, specifically a convolutional neural bidirectional recurrent network, to generate impulse responses from textual descriptions, incorporating BERT embedding technology and hyperparameters like Top-K and Top-P for filtering and temperature adjustment.
Enables accurate and controllable generation of impulse responses, achieving high performance and sound quality, applicable to digital amplifier simulation and virtual sound design.
Smart Images
Figure TWG2TA001073958_001 
Figure TWG2TA001073958_002 
Figure TWG2TA001073958_003
Abstract
Description
Technical Field
[0001] A method for generating impulse response, particularly a method for generating impulse response for a guitar amplifier. Prior Technology
[0002] Generative artificial intelligence-based text prompt generation technology has been widely applied in natural language processing and audio generation fields, such as music generation, speech synthesis, and sound simulation. However, current technologies for generating speaker impulse response (IR) still mainly rely on traditional measurement and recording methods, making it difficult to achieve rapid response and accurate simulation for personalized needs. Summary of the Invention
[0003] To improve upon current techniques for generating speaker impulse responses, which still primarily rely on traditional measurement and recording methods, this invention provides a method for generating impulse responses, comprising the following steps: Input an impulse response specification as text, generate a high-dimensional semantic vector using an embedding model, and then input the vector into an impulse response model; and The impulse response model generates one of the corresponding impulse responses based on the impulse response specification.
[0004] The impulse response model further introduces a temperature parameter to adapt to the adjustment and a hyperparameter for the filtering mechanism.
[0005] Preferably, the hyperparameter includes Top-K and / or Top-P.
[0006] The training steps for this impulse response model include: The raw data of several impulse responses are structured to form a data table. For each missing portion of the original data for that impulse response in the data table, a unified label is used, and the prediction space is set based on the existing attribute information; A one-dimensional convolutional neural network model was used to perform multi-label classification and feature completion on the data table, resulting in a complete data table; and Based on this complete data table, a convolutional neural bidirectional recurrent network architecture was used for training to obtain the impulse response model.
[0007] The data table contains at least the speaker type, loudspeaker type, configuration, open / closed, timbre, and / or microphone type for each raw data of the impulse response.
[0008] The Top-K filtering hyperparameter is performed by the following equation (1): …Formula (1).
[0009] The Top-p hyperparameter filtering is performed using the following equation (2): …Formula (2).
[0010] The mathematical formula for adjusting the temperature parameter is as follows (3): …Formula (3).
[0011] In response to the issue of missing attributes in some data in the data table, a unified marking method was adopted, marking missing values as "unknown".
[0012] In this convolutional neural network, the convolutional neural network layer extracts local information from the embedded features, the bidirectional recurrent network layer learns long-term dependencies and sequence correlations, and finally outputs a waveform data of the impulse response of a fixed length through a fully connected layer.
[0013] As can be seen from the above description, the present invention has the following beneficial effects and advantages:
[0014] 1. This invention is a technology for generating guitar amplifier impulse responses based on textual prompts. It combines deep learning models and BERT embedding technology to achieve accurate generation of amplifier IR from textual descriptions. By designing a convolutional neural bidirectional recurrent network generation model, it simulates diverse amplifier features and dynamically adjusts the similarity and diversity of the generated results. This technology can be effectively applied to digital amplifier simulation systems, virtual sound design, and digital guitar effects, possessing high practicality and market potential.
[0015] 2. The technology for generating impulse response audio based on text prompts provided in this invention can be used to simulate the frequency response characteristics of a virtual guitar amplifier. Combining natural language processing and generative deep learning models, feature vectors are extracted from text descriptions to generate highly accurate and controllable impulse response audio. A preferred embodiment of this invention uses a BERT model for text embedding and designs a one-dimensional convolutional neural network model to fill in missing labels in the dataset, constructing a complete and structured dataset. Furthermore, the model training employs K-Fold cross-validation to further improve generalization ability and robustness. Test results show that the generated impulse response audio achieves high levels of performance and sound quality, and a balance between similarity and diversity can be achieved by adjusting temperature parameters. This demonstrates that this invention not only expands the application of artificial intelligence generation technology in the field of amplifier simulation but also provides a new approach for the design and development of digital guitar effects. Simple Explanation of the Diagram
[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are merely some examples or embodiments of the present invention and are not intended to absolutely limit the technical scope of the present invention. Unless obvious from the context or otherwise stated, the same reference numerals in the drawings represent the same structures or operations. Wherein: Figure 1 is a flowchart of the steps of the impulse response generation method of the present invention. Figure 2 is a flowchart of step S1 in the impulse response generation method of the present invention. Figure 3 is a schematic flowchart of another step S1 in the impulse response generation method of the present invention. Figure 4 is a flowchart of step S2 in the impulse response generation method of the present invention. Figure 5 is a schematic flowchart of a preferred embodiment of the impulse response generation method of the present invention. Implementation
[0017] The present invention will be described in detail below with reference to several preferred embodiments. The accompanying drawings are merely some exemplary representations or embodiments of the present invention. For those skilled in the art, the present invention can be applied to other similar situations based on these drawings without any further effort.
[0018] The terms “system,” “apparatus,” “unit,” and / or “module” used in this invention are to distinguish different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they may be replaced by other expressions. As illustrated in this invention, unless the context clearly indicates otherwise, words such as “a,” “an,” “an,” and / or “the” do not specifically refer to the singular and may also include the plural. Generally, the terms “comprising” and “including” only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list; the method or apparatus may also include other steps or elements.
[0019] This invention uses flowcharts to illustrate the operations performed by the system according to embodiments of the invention. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, the steps can be processed in reverse order or simultaneously. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.
[0020] <Methods for generating impulse responses>
[0021] Please refer to Figure 1, which is a flowchart of the steps of the impulse response generation method of the present invention, the steps of which include:
[0022] As shown in Figure 2, step S1: Database establishment and model training
[0023] Step S1-1 Data Establishment: Several impulse response raw data are structured to form a data table; this data table should at least include the cabinet type, speaker type, configuration, open / closed, tone, and / or microphone type for each impulse response raw data, as shown in Figure 3;
[0024] Step S1-2: For each missing portion of the original impulse response data in the data table, mark it uniformly, and set the prediction space based on existing attribute information;
[0025] Steps S1-3: Use a one-dimensional convolutional neural network (1D-CNN) model to perform multi-label classification and feature completion on the data table to obtain a complete data table;
[0026] Steps S1-4: Based on the complete data table, a convolutional neural bidirectional recurrent network (CNN-BiGRU) is used to generate an architecture for training, resulting in an impulse response model;
[0027] Please refer to step S2 in Figures 1 and 4: Generate impulse response based on text prompts.
[0028] Step S2-1: Input an impulse response specification as text, generate a high-dimensional semantic vector using an embedding model, and then input the vector into the impulse response model;
[0029] Step S2-2: The impulse response model generates one of the corresponding impulse responses based on the impulse response specification.
[0030] In this bidirectional recurrent neural network, the convolutional neural network (CNN) layer extracts local information from the embedded features, the bidirectional recurrent network (BiGRU) layer learns long-term dependencies and sequence correlations, and finally outputs a waveform data of the impulse response of fixed length through a fully connected layer.
[0031] Step S6 (optional): To further improve the audio diversity and flexibility of generating the impulse response, the impulse response model in step S2-2 can introduce a temperature parameter for adaptation and a hyperparameter for filtering mechanism. The hyperparameter includes Top-K and / or Top-P (also known as cumulative probability threshold or core sampling) to achieve fine control over the generation of the impulse response.
[0032] The Top-K filtering hyperparameter is performed by the following equation (1): …Formula (1);
[0033] The hyperparameter Top-p filtering is performed by the following equation (2): …Formula (2);
[0034] The mathematical formula for adjusting the temperature parameter is as follows (3): …Formula (3);
[0035] The symbols in equations (1) to (3) are defined as follows in Table 1:
[0036] Table 1. symbol definition Represented as candidate options, for example, the i-th candidate option after sorting (from highest to lowest), and so on. The original output scores of the BERT model The BERT model's original output score for the i-th candidate after sorting, compared with... The effect is the same, only represented by a numerical index. Temperature parameters control the smoothness of the distribution. Adjusted probability distribution When the cumulative probability When p is less than or equal to p, the logits value is retained, and the probability distribution is calculated using softmax. This involves first exponentially operating on the logits and then normalizing it to a probability. Top-k candidates For the candidates After sorting, select the top k candidates with the highest probability. Top-p Top-p is a probability threshold that retains the range where the total probability does not exceed p.
[0037] The aforementioned impulse response generation method of the present invention can be computed directly in the cloud or in hardware, including but not limited to desktop computers, laptops, mobile phones or tablets.
[0038] < [Example 1]
[0039] Please refer to Figure 5. A preferred embodiment of the present invention takes the generation of a guitar amplifier impulse response as an example.
[0040] Step S1: Database Establishment and Model Training
[0041] Step S1-1: This embodiment uses four raw impulse response datasets from different sources and performs structured processing to ultimately form a unified data table containing 46,564 data entries;
[0042] Please refer to Tables 2-1 and 2-2 below. In the process of creating this data table, firstly, all the audio files of the impulse response raw data and their descriptive text are integrated into a table with multiple attribute columns. These attributes include at least the cabinet type, speaker type, configuration, open / closed, tone, and microphone type for each impulse response raw data.
[0043] Step S2: To address the issue of missing attributes in some data in the data table, a unified labeling method is adopted to mark missing values as "unknown", and the prediction space is set based on the existing attribute information.
[0044] Table 2-1 shows a preferred embodiment of this data table (only a portion and codes are presented). Cabinet Speaker Type A 412 A30 B 412 B20 C K412 C30
[0045] Table 2-2 (continued), a preferred embodiment of this data table (only a portion and codes are presented). Configuration Open(O) / Close(C) Tone Microphone 412 C Dark 313 412 C Tight 67 412 C Thick 906 112 C unknown 121+47 212 C Balanced 121 212 C Balanced 414 212 C Thick 57
[0046] Step S3: Using the one-dimensional convolutional neural network (1D-CNN) model, multi-label classification and feature completion are performed on the data table. Specifically, the CNN model extracts and classifies temporal features from each IR audio file, infers the values of unknown attributes, and automatically fills the data table with the results. It accurately predicts and fills in attributes such as "speaker type," "loudspeaker type," and "microphone type," thereby generating a structured dataset containing 46,564 data points. The one-dimensional convolutional neural network then classifies the original impulse response data and identifies key features. The input to the one-dimensional convolutional neural network is the original impulse response audio file. After multiple convolutions and feature extraction, it predicts and outputs the corresponding structured features. During this process, through the local feature extraction capability of the one-dimensional convolutional neural network, the model can effectively identify key features in the audio, further improving the data completeness and accuracy of the subsequent generated models.
[0047] Steps S1-4: Based on the complete data table, a convolutional neural bidirectional recurrent network (CNN-BiGRU) is used to generate an architecture for training, resulting in an impulse response model;
[0048] Step S2: Generate impulse response based on text prompts
[0049] Step S2-1: Input an impulse response specification in text, generate a high-dimensional semantic vector using an embedding model, and then input the impulse response model; in this embodiment, the input impulse response specification is text input (Text Prompt) = "I need an impulse response profile containing Mrshall60AV speakers, Celestion G12 Vintage speakers, 4*12 configuration and AKG C414 microphone".
[0050] Step S2-2: The impulse response model generates one of the corresponding impulse responses based on the impulse response specification.
[0051] In the foregoing description, some embodiments use numbers describing the quantity of components or attributes. It should be understood that such numbers used in the description of embodiments are sometimes modified by the terms "approximately," "approximately," or "generally." Unless otherwise stated, "approximately," "approximately," or "generally" indicates that the numbers are allowed to vary by ±20%. Accordingly, in some embodiments, the numerical parameters used in the specification and claims are approximate values, which may be changed depending on the characteristics required by individual embodiments. In some embodiments, numerical parameters should take into account specified significant digits and employ a general method of digit reservation. Although the numerical ranges and parameters used to confirm their breadth of range in some embodiments of the invention are approximate values, in specific embodiments, such values are set as precisely as feasible.
[0052] Finally, it should be understood that the embodiments described in this invention are merely illustrative of the principles of the invention. Other modifications may also fall within the scope of this invention. Therefore, alternative configurations of the embodiments of this invention are considered as examples and not limitations, and are regarded as consistent with the teachings of this invention. Accordingly, the embodiments of this invention are not limited to those explicitly described and illustrated herein.
[0053] S1, S2, S1-1~S2-2: Steps
Claims
1. A method for generating an impulse response, comprising the steps of: inputting an impulse response specification as text, generating a high-dimensional semantic vector using an embedding model, and then inputting the vector into an impulse response model; and the impulse response model generating a corresponding impulse response based on the impulse response specification; the impulse response model incorporating a temperature parameter for adaptive adjustment and a hyperparameter for filtering.
2. The impulse response generation method as described in claim 1, wherein: the hyperparameter includes Top-K and / or Top-P.
3. The impulse response generation method as described in claim 1, wherein: the training steps of the impulse response model include: structuring several raw impulse response data to form a data table; uniformly labeling the missing parts of each raw impulse response data in the data table, and setting the prediction space based on existing attribute information; using a one-dimensional convolutional neural network model to perform multi-label classification and feature completion of the missing parts of the data table to obtain a complete data table; and training the impulse response model based on the complete data table using a convolutional neural bidirectional recurrent network generation architecture.
4. The impulse response generation method as described in claim 3, wherein: the data table contains at least the speaker type, loudspeaker type, configuration, open / closed, timbre, and / or microphone type for each of the raw impulse response data.
5. The impulse response generation method as described in Request 1, wherein: the hyperparameter Top-K filtering is performed by the following equation (1): … Equation (1), where: is the i-th candidate after sorting (from high to low); is the original output score of the BERT model; and Top-k candidates are the top k candidates with the highest probability after sorting the logits() of the candidates.
6. The impulse response generation method as described in Request 1, wherein: the hyperparameter Top-p filtering is performed by the following equation (2): …Equation (2), where: is the i-th candidate after sorting (from high to low); is the j-th candidate after sorting (from high to low); P() retains the logits value when the cumulative probability is less than or equal to p, and the probability distribution is calculated through softmax, that is, the logits are first exponentially operated and then normalized to the probability; is the original output score generated by the BERT model for the i-th candidate after sorting; and Top-p and p are probability thresholds, retaining the range where the total probability does not exceed p.
7. The impulse response generation method as described in Request 1, wherein: the mathematical formula for adjusting the temperature parameter is as follows (3): … Formula (3), wherein: is the i-th candidate after sorting (from high to low); T is the temperature parameter; is the original output score generated by the BERT model for the i-th candidate after sorting, which is the same as the effect of logits(), but is represented by a numerical index; and is the probability distribution of the candidate after temperature T adjustment.
8. The impulse response generation method as described in request item 3, wherein: for the problem of missing attributes in some data in the data table, a unified marking method is adopted to mark the missing values as "unknown".
9. The impulse response generation method as described in claim 3, wherein: the convolutional neural network layer of the bidirectional recurrent neural network extracts local information from the embedded features, the bidirectional recurrent network layer learns long-term dependencies and sequence correlations, and finally outputs a waveform data of the impulse response of fixed length through a fully connected layer.