技术领域Technical Field
本发明涉及语音识别技术领域，具体为一种语音内容识别方法及系统。The present invention relates to the field of speech recognition technology, and in particular to a speech content recognition method and system.
背景技术Background Art
现有的语音内容识别技术已广泛应用于智能助手、自动翻译、语音控制系统等领域，极大地提高了人机交互的便捷性。然而，这些技术通常面临准确性和实时性的挑战。尤其是在嘈杂环境下，背景噪声容易影响语音信号的质量，导致识别错误。此外，现有系统对于不同用户的口音、语速和语言习惯的适应性不足，限制了语音识别技术的普及和应用范围。Existing speech content recognition technologies have been widely used in the fields of intelligent assistants, automatic translation, voice control systems, etc., greatly improving the convenience of human-computer interaction. However, these technologies usually face challenges in accuracy and real-time performance. Especially in noisy environments, background noise can easily affect the quality of speech signals, leading to recognition errors. In addition, the existing systems are not adaptable enough to the accents, speaking speeds, and language habits of different users, which limits the popularity and application scope of speech recognition technology.
尽管已有多种改进方法被提出，如采用更复杂的声学模型和语言模型来提高识别准确率，但这些方法往往需要大量的计算资源，影响了系统的实时响应能力。同时，现有技术在处理特定行业的专业术语或用户个性化语言习惯时，常常表现出识别能力不足。此外，现有系统在用户身份验证方面也存在缺陷，难以有效区分真实用户和潜在的欺诈者，这在需要高安全性的应用场景中尤为突出。Although a variety of improvement methods have been proposed, such as using more complex acoustic models and language models to improve recognition accuracy, these methods often require a lot of computing resources, affecting the system's real-time response capabilities. At the same time, existing technologies often show insufficient recognition capabilities when dealing with professional terms in specific industries or users' personalized language habits. In addition, existing systems also have defects in user identity authentication, making it difficult to effectively distinguish between real users and potential fraudsters, which is particularly prominent in application scenarios that require high security.
为了解决上述缺陷，现提供一种技术方案。In order to solve the above defects, a technical solution is now provided.
发明内容Summary of the invention
本发明的目的在于现有语音识别技术在准确性、实时性、个性化适应性以及安全性方面的不足，而提出一种语音内容识别方法及系统。The purpose of the present invention is to address the deficiencies of existing speech recognition technology in terms of accuracy, real-time performance, personalized adaptability and security, and to propose a speech content recognition method and system.
本发明的目的可以通过以下技术方案实现：The purpose of the present invention can be achieved through the following technical solutions:
一种语音内容识别方法，包括以下步骤：A method for speech content recognition comprises the following steps:
S1、用户确认：通过语音内容确认用户，具体过程如下：S1. User confirmation: Confirm the user through voice content. The specific process is as follows:
用户首次使用时，朗读预定义文本以收集语音样本，并从中提取音高、音调特征创建声音模板；When the user uses it for the first time, the predefined text is read aloud to collect voice samples, and pitch and tone features are extracted from them to create a sound template;
对采集的语音进行降噪和声音增强，使用傅里叶变换和MFCC技术提取语音特征，将提取的语音特征和声音模板存储并与用户账户关联；Perform noise reduction and sound enhancement on the collected speech, extract speech features using Fourier transform and MFCC technology, store the extracted speech features and sound templates and associate them with the user account;
在身份验证时，实时采集并处理用户的语音，提取其特征，利用DTW算法或GMM技术，将实时语音特征与声音模板匹配；During identity authentication, the user's voice is collected and processed in real time, its features are extracted, and the real-time voice features are matched with the sound template using the DTW algorithm or GMM technology;
设定阈值以判断匹配结果，低于阈值则确认用户身份，实施活体检测机制，确保采集的是实时语音，并结合生物特征或密码，增强安全性；Set a threshold to determine the matching result. If the result is below the threshold, the user's identity is confirmed. Implement a liveness detection mechanism to ensure that real-time voice is collected, and combine it with biometrics or passwords to enhance security.
匹配失败时，重新匹配或作为新用户录入，定期更新声音模板以适应用户声音变化，提供用户界面反馈匹配结果，并管理FAR和FRR以平衡安全性与便利性；When matching fails, rematch or register as a new user, regularly update voice templates to adapt to user voice changes, provide user interface feedback on matching results, and manage FAR and FRR to balance security and convenience;
S2、用户资料存储：存储不同用户的声音模板数据；S2, user data storage: storing voice template data of different users;
S3、语音内容识别：对用户的语音进行实时的内容识别；S3, voice content recognition: real-time content recognition of the user's voice;
S4、个性化调整：根据用户的语音习惯进行内容识别的调整；S4, Personalized adjustment: adjust content recognition according to the user's voice habits;
S5、反馈优化：根据用户的反馈或系统判断的修正进行不断优化识别内容。S5. Feedback optimization: Continuously optimize the recognition content based on user feedback or corrections made by the system.
进一步的，所述S1中实施活体检测机制，确保采集的是实时语音的具体步骤如下：Furthermore, the specific steps of implementing the liveness detection mechanism in S1 to ensure that the real-time voice is collected are as follows:
选择活体检测方法，包括唇动检测、语音节奏分析或随机挑战响应，将选定的活体检测方法集成到用户确认流程中；Select a liveness detection method, including lip movement detection, speech rhythm analysis, or random challenge response, and integrate the selected liveness detection method into the user confirmation process;
在用户验证时，同时采集语音和面部视频数据，通过视频分析唇动与语音的同步性；During user verification, voice and facial video data are collected simultaneously, and the synchronization between lip movement and voice is analyzed through video;
评估语音的节奏、强度和语调的自然性；Assess the naturalness of speech rhythm, intensity, and intonation;
要求用户响应随机生成的指令或问题；Require users to respond to randomly generated instructions or questions;
检测语音中的噪声和干扰，辨识非实时语音，分析检测结果，确定是否为实时语音，如检测失败，提示用户并允许重试验证。Detect noise and interference in speech, identify non-real-time speech, analyze the detection results, and determine whether it is real-time speech. If the detection fails, prompt the user and allow retest verification.
进一步的，所述S2的具体操作步骤如下：Furthermore, the specific operation steps of S2 are as follows:
引导用户朗读预设文本或回答问题，收集用户的语音，对样本进行去噪、归一化和分割的预处理操作；Guide users to read preset text or answer questions, collect user voice, and perform pre-processing operations such as denoising, normalization, and segmentation on samples;
使用MFCC技术提取声音模块中的语音特征，分析得到用户的对应的语音属性，根据分析结果创建包含关键特征的声音模板；Use MFCC technology to extract speech features in the sound module, analyze the user's corresponding speech attributes, and create a sound template containing key features based on the analysis results;
将声音模板和特征数据进行存储，并与用户账号关联，建立索引用于快速检索用户语音数据，同时定期备份数据，防止丢失或损坏；The voice templates and feature data are stored and associated with the user account. An index is created for quick retrieval of user voice data. The data is backed up regularly to prevent loss or damage.
允许用户定期更新样本，适应声音变化，定期清理和维护数据，确保信息的时效性和准确性。Allow users to regularly update samples, adapt to sound changes, and regularly clean and maintain data to ensure the timeliness and accuracy of information.
进一步的，所述S3的具体操作步骤如下：Furthermore, the specific operation steps of S3 are as follows:
捕获并记录用户的实时语音，执行去噪、回声消除和增益控制的操作优化语音信号，从优化后的信号中提取MFCC、音高和音调的语音特征；Capture and record the user's real-time voice, perform denoising, echo cancellation and gain control operations to optimize the voice signal, and extract MFCC, pitch and tone voice features from the optimized signal;
应用声学模型将特征映射到声学单元，结合语言模型提升识别的上下文准确性；Apply acoustic models to map features to acoustic units, and combine with language models to improve the accuracy of context recognition;
对模型输出进行解码，生成候选词序列，分析并选择最符合语境的词或短语；Decode the model output, generate candidate word sequences, analyze and select the words or phrases that best fit the context;
将识别的文本输出供用户使用或进一步处理。Output the recognized text for user consumption or further processing.
进一步的，所述S3中分析并选择最符合语境的词或短语的具体操作步骤如下：Furthermore, the specific operation steps of analyzing and selecting the words or phrases that best fit the context in S3 are as follows:
对每个候选词的在当前上下文的语法和语义合适度进行综合分析，通过分析候选词的合适评价参数，其中合适评价参数包括：A comprehensive analysis is performed on the grammatical and semantic suitability of each candidate word in the current context by analyzing the appropriate evaluation parameters of the candidate word, where the appropriate evaluation parameters include:
语法合适度：评估候选词是否符合当前句子的语法结构，将评估结果通过语法合适度评分进行量化，并记为语法值，以此语法值作为衡量语法合适度的标准；Grammatical suitability: Evaluate whether the candidate word conforms to the grammatical structure of the current sentence. The evaluation result is quantified through the grammatical suitability score and recorded as the grammatical value, which is used as the standard for measuring grammatical suitability.
语义合适度：评估候选词是否在语义上与上下文相匹配，包括词义的一致性和逻辑性，将评估结果通过评分进行量化，并记为语义值，以此语义值作为衡量语义合适度的标准；Semantic fit: Evaluate whether the candidate word matches the context semantically, including the consistency and logic of the word meaning. The evaluation result is quantified through scoring and recorded as a semantic value, which is used as a standard to measure semantic fit;
概率评分：通过基于n-gram模型或神经网络语言模型为候选词序列分配概率评分；Probability scoring: assigning probability scores to candidate word sequences based on n-gram models or neural network language models;
词频统计：分析候选词在上下文中的出现频率；Word frequency statistics: analyze the frequency of occurrence of candidate words in the context;
词向量相似度：使用词嵌入技术评估候选词与上下文中其他词的向量空间距离；Word vector similarity: Use word embedding technology to evaluate the vector space distance between the candidate word and other words in the context;
依赖关系评分：在句法分析的基础上，评估候选词与上下文中其他词的句法依赖关系，并通过赋予依赖关系评分进行量化；Dependency scoring: Based on syntactic analysis, the syntactic dependency between the candidate word and other words in the context is evaluated and quantified by assigning a dependency score;
再分别将得到的语法值、语义值、概率评分、出现频率、向量空间距离及依赖关系评分标定为YF、YY、GP、CP、KJ及YP，归一化处理后代入以下公式：以得到合评值HPZ，式中分别为语法值、语义值、概率评分、出现频率、向量空间距离及依赖关系评分的预设权重系数，并以得到的合评值作为衡量目前分析的候选词在当前上下文的语法和语义合适度的综合评价标准；The obtained grammatical value, semantic value, probability score, occurrence frequency, vector space distance and dependency score are calibrated as YF, YY, GP, CP, KJ and YP respectively, and after normalization, they are entered into the following formula: To obtain the combined evaluation value HPZ, The preset weight coefficients are respectively the grammatical value, semantic value, probability score, occurrence frequency, vector space distance and dependency score, and the obtained combined score is used as a comprehensive evaluation criterion for measuring the grammatical and semantic suitability of the candidate word currently analyzed in the current context;
将所有候选词得到的合评值按照大小进行排序，选取合评值最大的候选词作为最优项，同时在用户界面展示最优项的分析结果，并提供选项供用户确认或更正。The combined evaluation values of all candidate words are sorted in order, and the candidate word with the largest combined evaluation value is selected as the optimal option. The analysis result of the optimal option is displayed on the user interface, and options are provided for the user to confirm or correct it.
进一步的，所述S4的具体操作步骤如下：Furthermore, the specific operation steps of S4 are as follows:
确认语音用户身份，并提取存储该用户的语音特征，利用用户数据更新个性化声学和语言模型；Confirm the identity of the voice user, extract and store the user's voice features, and use the user data to update the personalized acoustic and language models;
分析识别结果，识别潜在的错误或不确定性，结合用户上下文和历史数据，提高系统对用户语言习惯的理解；Analyze recognition results, identify potential errors or uncertainties, and combine user context and historical data to improve the system's understanding of user language habits;
创建包含用户特定术语和缩写的个性化词典，调整敏感度及识别阈值参数以适应用户特征；Create a personalized dictionary containing user-specific terms and abbreviations, and adjust sensitivity and recognition threshold parameters to suit user characteristics;
根据用户反馈修正错误，并通过自适应学习算法自动改进，结合用户的文本输入习惯等数据源，提供个性化服务；Correct errors based on user feedback and automatically improve through adaptive learning algorithms, combining data sources such as user text input habits to provide personalized services;
续监控系统性能，确保优化措施有效并进行及时调整。Continue to monitor system performance to ensure optimization measures are effective and make timely adjustments.
进一步的，所述的S4中分析识别结果，识别潜在的错误或不确定性具体操作步骤如下：Furthermore, the specific steps of analyzing the recognition results in S4 and identifying potential errors or uncertainties are as follows:
收集识别文本和置信度评分，评估置信度以识别不确定性或错误，分析错误类型，包括替换、插入或删除错误；Collect recognition text and confidence scores, evaluate confidence to identify uncertainties or errors, and analyze error types, including substitution, insertion, or deletion errors;
验证识别结果的语法、语义和逻辑是否与上下文相符；Verify whether the syntax, semantics and logic of the recognition results are consistent with the context;
允许用户通过界面标记错误，并将反馈用于分析，记录所有错误和不确定性实例，用于分析和模型训练；Allow users to mark errors through the interface and use the feedback for analysis, recording all errors and uncertainty instances for analysis and model training;
确定错误背后的原因，包括环境干扰或模型不匹配；Determine the reasons behind the errors, including environmental interference or model mismatch;
比较识别结果与候选词列表，评估替代选项，根据错误分析更新个性化词典和语法规则，优化识别准确性。Compare recognition results with candidate word lists, evaluate alternative options, and update personalized dictionaries and grammar rules based on error analysis to optimize recognition accuracy.
进一步的，所述的S5的具体操作步骤如下：Furthermore, the specific operation steps of S5 are as follows:
通过用户界面获取用户对识别结果的评价和错误报告，自动记录未确认和标记的识别错误，并进行标注；Obtain user comments and error reports on recognition results through the user interface, automatically record unconfirmed and marked recognition errors, and annotate them;
析错误的原因，包括环境、模型不足或口音差异，利用反馈数据更新声学和语言模型，优化识别算法，并根据反馈调整个性化词典和语法规则；Analyze the causes of errors, including environment, model deficiencies, or accent differences, use feedback data to update acoustic and language models, optimize recognition algorithms, and adjust personalized dictionaries and grammar rules based on feedback;
通过自适应机制自动调整系统以适应用户语音习惯，确保用户反馈被及时处理并用于系统改进；Automatically adjust the system to adapt to user speech habits through adaptive mechanisms, ensuring that user feedback is processed in a timely manner and used for system improvement;
创建积分或奖励系统以鼓励用户反馈，分析用户需求和动机，确定最优激励措施。Create a points or reward system to encourage user feedback, analyze user needs and motivations, and determine the optimal incentives.
进一步的，所述的S5中分析用户需求和动机，确定最优激励措施的具体操作步骤如下：Furthermore, the specific steps of analyzing user needs and motivations and determining the optimal incentive measures in S5 are as follows:
采用在线调查、AI电话访问或焦点论坛方法进行用户调研，选取用户样本进行调研，列出激励措施选项，包括积分、折扣及开放性问题；Conduct user research using online surveys, AI telephone interviews, or focus forums, select user samples for research, and list incentive options, including points, discounts, and open questions;
让用户对激励措施进行评分和排名，收集用户的偏好信息，通过多种渠道分发问卷，并确保回收数据的准确性；Allow users to rate and rank incentives, collect information on user preferences, distribute questionnaires through multiple channels, and ensure the accuracy of returned data;
整理和分析数据，使用统计方法确定用户偏好，分析不同用户群体的偏好差异，将分析结果整理成报告，并提出推荐措施，根据调研结果，选择激励措施进行预设规模的测试并进行效果评估，判断效果是否达标，若达标，确定以该激励措施作为最终选择。Organize and analyze data, use statistical methods to determine user preferences, analyze the differences in preferences among different user groups, compile the analysis results into a report, and make recommendations. Based on the survey results, select incentive measures for testing on a preset scale and conduct an effectiveness evaluation to determine whether the results meet the standards. If so, determine that the incentive measure will be the final choice.
一种语音内容识别系统，包括：A speech content recognition system, comprising:
用户资料存储模块，用于存储不同用户的声音模板数据；User data storage module, used to store voice template data of different users;
语音内容识别模块， 用于对用户的语音进行实时的内容识别；Voice content recognition module, used to perform real-time content recognition on the user's voice;
个性化优化模块，用于根据用户的语音习惯进行内容识别的优化；Personalized optimization module, used to optimize content recognition based on user's voice habits;
反馈模块，用于根据用户的反馈或系统判断的修正进行不断优化识别内容。The feedback module is used to continuously optimize the recognition content based on user feedback or corrections made by the system.
与现有技术相比，本发明的有益效果是：Compared with the prior art, the present invention has the following beneficial effects:
本发明，通过集成先进的语音信号处理技术和活体检测机制，显著提高了语音识别的准确性和安全性，用户确认模块采用声音模板匹配和活体检测技术，如唇动检测和语音节奏分析，有效区分了真实用户和潜在的欺诈者，确保了只有用户实时的语音才能被系统接受，通过定期更新声音模板以适应用户声音的变化，增强了系统的适应性和长期稳定性；The present invention significantly improves the accuracy and security of speech recognition by integrating advanced speech signal processing technology and liveness detection mechanism. The user confirmation module adopts sound template matching and liveness detection technology, such as lip movement detection and speech rhythm analysis, to effectively distinguish real users from potential fraudsters, ensuring that only the user's real-time speech can be accepted by the system. By regularly updating the sound template to adapt to the changes in the user's voice, the adaptability and long-term stability of the system are enhanced.
本发明，个性化调整模块通过分析用户的语音习惯，包括口音、语速和常用词汇，实现了对声学和语言模型的个性化更新，结合用户反馈和自适应学习算法，系统能够自动调整识别参数，如敏感度和识别阈值，以适应不同用户的特定需求，这种个性化的服务不仅提升了用户体验，也增强了系统对各种语音输入的鲁棒性；The present invention realizes personalized update of acoustic and language models by analyzing the user's voice habits, including accent, speaking speed and common vocabulary. Combining user feedback and adaptive learning algorithm, the system can automatically adjust recognition parameters, such as sensitivity and recognition threshold, to meet the specific needs of different users. This personalized service not only improves the user experience, but also enhances the robustness of the system to various voice inputs.
本发明，反馈优化模块通过用户界面获取用户评价和错误报告，自动记录并分析系统识别错误，从而不断优化识别算法和个性化设置，通过设计激励机制鼓励用户提供反馈，建立了一个积极的用户参与环境，促进了系统的持续改进和性能提升，通过这种反馈循环，系统能够及时响应用户需求，快速迭代更新，以满足不断变化的市场需求。In the present invention, the feedback optimization module obtains user comments and error reports through the user interface, automatically records and analyzes system recognition errors, thereby continuously optimizing the recognition algorithm and personalized settings, and encourages users to provide feedback by designing an incentive mechanism, thereby establishing a positive user participation environment and promoting continuous improvement and performance enhancement of the system. Through this feedback loop, the system can respond to user needs in a timely manner and quickly iterate and update to meet the ever-changing market demands.
附图说明BRIEF DESCRIPTION OF THE DRAWINGS
为了便于本领域技术人员理解，下面结合附图对本发明作进一步的说明；In order to facilitate understanding by those skilled in the art, the present invention is further described below in conjunction with the accompanying drawings;
图1为本发明的方法流程图。FIG1 is a flow chart of the method of the present invention.
具体实施方式DETAILED DESCRIPTION
下面将结合实施例对本发明的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本发明一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其它实施例，都属于本发明保护的范围。The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
应当理解，本披露的说明书和权利要求书中使用的术语“包括”和 “包含”指示所描述特征、整体、步骤、操作、元素和/或组件的存在，但并不排除一个或多个其它特征、整体、步骤、操作、元素、组件和/或其集合的存在或添加。It should be understood that the terms "include" and "comprising" used in the specification and claims of the present disclosure indicate the presence of described features, integers, steps, operations, elements and/or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and/or collections thereof.
还应当理解，在此本披露说明书中所使用的术语仅仅是出于描述特定实施例的目的，而并不意在限定本披露。如在本披露说明书和权利要求书中所使用的那样，除非上下文清楚地指明其它情况，否则单数形式的“一”、“一个”及“该”意在包括复数形式。还应当进一步理解，在本披露说明书和权利要求书中使用的术语“和/ 或”是指相关联列出的项中的一个或多个的任何组合以及所有可能组合，并且包括这些组合。It should also be understood that the terms used in this disclosure are only for the purpose of describing specific embodiments and are not intended to limit the disclosure. As used in this disclosure and claims, the singular forms of "a", "an", and "the" are intended to include the plural forms unless the context clearly indicates otherwise. It should also be further understood that the term "and/or" used in this disclosure and claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.
如图1所示，一种语音内容识别方法，包括以下步骤：As shown in FIG1 , a method for speech content recognition includes the following steps:
用户确认：User acknowledgement:
在用户首次使用系统时，要求用户朗读一组预定义的句子或数字序列，以收集用户的语音样本，从这些样本中提取语音特征，如音高、音调、节奏、音色等，并创建用户的声音模板；对用户的语音样本进行预处理，包括降噪、声音增强等，以提高特征提取的准确性，使用语音信号处理技术，如傅里叶变换、梅尔频率倒谱系数（MFCC）等，提取关键的语音特征；When a user uses the system for the first time, the user is required to read a set of predefined sentences or digital sequences to collect the user's voice samples, extract voice features such as pitch, tone, rhythm, timbre, etc. from these samples, and create the user's voice template; pre-process the user's voice samples, including noise reduction, sound enhancement, etc., to improve the accuracy of feature extraction, and use voice signal processing technologies such as Fourier transform and Mel Frequency Cepstral Coefficient (MFCC) to extract key voice features;
其中傅里叶变换是一种数学变换，用于将信号（在语音识别中是语音信号）从时间域转换到频率域。这种转换帮助分析信号中不同频率成分的强度和分布，进而提取出对语音识别有用的特征，如音高和音色；The Fourier transform is a mathematical transformation used to convert a signal (in speech recognition, it is a speech signal) from the time domain to the frequency domain. This transformation helps analyze the intensity and distribution of different frequency components in the signal, and then extract features useful for speech recognition, such as pitch and timbre;
梅尔频率倒谱系数（MFCC）是一种在语音处理中广泛使用的声学特征表示方法，它基于人耳的听觉感知特性，通过一系列步骤从语音信号的功率谱中提取出反映语音信号主要特征的系数。MFCC能够捕捉到语音信号的重要特征，如音高、音色等，并且对噪声和录音质量的变化具有一定的鲁棒性；Mel-frequency cepstral coefficient (MFCC) is an acoustic feature representation method widely used in speech processing. It is based on the auditory perception characteristics of the human ear and extracts coefficients reflecting the main characteristics of the speech signal from the power spectrum of the speech signal through a series of steps. MFCC can capture important features of speech signals, such as pitch, timbre, etc., and has a certain degree of robustness to changes in noise and recording quality;
提取的语音特征和声音模板存储在用户资料存储模块中，与用户账户关联，当用户需要进行身份验证时，系统实时采集用户的语音输入，对实时采集的语音进行同样的预处理和特征提取步骤；使用声纹识别技术，将实时提取的语音特征与存储的声音模板进行匹配，计算特征之间的相似度或距离，如使用动态时间弯曲（DTW）算法或高斯混合模型（GMM）；The extracted voice features and sound templates are stored in the user data storage module and associated with the user account. When the user needs to authenticate, the system collects the user's voice input in real time and performs the same preprocessing and feature extraction steps on the real-time collected voice. Using voiceprint recognition technology, the real-time extracted voice features are matched with the stored sound templates, and the similarity or distance between the features is calculated, such as using the dynamic time warping (DTW) algorithm or the Gaussian mixture model (GMM).
其中使用动态时间弯曲（DTW）算法是一种用于度量两个时间序列之间相似度的算法，即使这两个序列在时间轴上存在伸缩或变形。在语音识别中，DTW常用于将实时采集的语音特征与预先录制的声音模板进行比较，通过计算两者之间的距离来评估它们的相似度；The Dynamic Time Warping (DTW) algorithm is an algorithm used to measure the similarity between two time series, even if the two series are stretched or deformed on the time axis. In speech recognition, DTW is often used to compare real-time collected speech features with pre-recorded sound templates, and evaluate their similarity by calculating the distance between the two;
高斯混合模型（GMM）是一种概率模型，用于表示具有多个子群体的数据集，其中每个子群体的数据分布可以用高斯分布（正态分布）来近似，在语音识别中，GMM可以用来建模声音模板的特征分布，进而用于评估实时语音特征与声音模板的匹配程度；Gaussian mixture model (GMM) is a probabilistic model used to represent a data set with multiple subgroups, where the data distribution of each subgroup can be approximated by Gaussian distribution (normal distribution). In speech recognition, GMM can be used to model the feature distribution of sound templates, and then used to evaluate the matching degree between real-time speech features and sound templates.
设置一个匹配阈值，当相似度或距离低于这个阈值时，认为匹配成功，确认用户身份；引入活体检测机制，确保采集到的是用户实时的语音，而非录音或合成语音，具体过程如下：A matching threshold is set. When the similarity or distance is lower than this threshold, the match is considered successful and the user's identity is confirmed. A liveness detection mechanism is introduced to ensure that the user's real-time voice is collected, rather than recorded or synthesized voice. The specific process is as follows:
确定使用哪种类型的活体检测技术，如唇动检测、语音节奏分析、随机挑战响应等，将选定的活体检测技术集成到用户确认模块中，确保在进行声音模板匹配前进行活体检测；在用户进行身份验证时，实时采集用户的语音输入，当采用唇动检测时，使用摄像头捕获用户的面部视频，分析用户的唇动与语音同步性，确保语音与唇动一致；Determine which type of liveness detection technology to use, such as lip movement detection, voice rhythm analysis, random challenge response, etc., and integrate the selected liveness detection technology into the user confirmation module to ensure liveness detection before performing sound template matching; when the user authenticates the identity, collect the user's voice input in real time. When lip movement detection is used, use a camera to capture the user's facial video, analyze the synchronization of the user's lip movement and voice, and ensure that the voice and lip movement are consistent;
当采用语音节奏分析时，分析用户语音的节奏、强度和语调变化，检测是否符合真人发声的自然模式；当采用随机挑战响应时，要求用户根据系统随机生成的指令或问题进行响应，以检测是否为真人实时反应；分析语音信号中的噪声和干扰模式，以识别录音重放或合成语音可能性，当可能性超过阈值时，则判断为录音重放或合成语音；当判断活体检测失败时，提示用户重新进行验证，并提供失败的原因。When using speech rhythm analysis, the rhythm, intensity and intonation changes of the user's voice are analyzed to detect whether it conforms to the natural pattern of real-life voice. When using random challenge response, the user is required to respond according to the instructions or questions randomly generated by the system to detect whether it is a real-time response. The noise and interference patterns in the voice signal are analyzed to identify the possibility of recorded playback or synthesized speech. When the possibility exceeds the threshold, it is judged to be recorded playback or synthesized speech. When it is judged that the liveness detection has failed, the user is prompted to re-verify and the reason for the failure is provided.
同时结合其他生物特征或传统密码进行多因素验证，以提高安全性；如果系统判断匹配失败或用户反馈识别错误，重新进行匹配或将该用户以新用户的身份重新录入；根据用户的使用习惯和声音变化，定期更新声音模板，使系统能够适应用户的变化；提供用户界面反馈，告知用户匹配结果，如成功、失败或需要重新输入，设计错误拒绝率（FAR）和错误接受率（FRR）的管理策略，平衡安全性和便利性。At the same time, multi-factor verification is performed in combination with other biometric features or traditional passwords to improve security. If the system determines that the match has failed or the user reports an identification error, the match is repeated or the user is re-entered as a new user. According to the user's usage habits and voice changes, the voice template is updated regularly so that the system can adapt to user changes. User interface feedback is provided to inform the user of the matching results, such as success, failure or the need for re-entry, and management strategies for the false rejection rate (FAR) and false acceptance rate (FRR) are designed to balance security and convenience.
用户资料存储：User data storage:
录入用户的注册信息，并进行语音样本的采集，具体通过引导用户录入预设的特定文本或回答特定问题，以收集用户的语音样本；对采集的语音样本进行预处理，包括去噪、归一化、分割等步骤，使用语音信号处理技术提取语音样本的特征，如梅尔频率倒谱系数（MFCC）、音高、音调等，对提取的语音特征进行分析，识别出用户独特的语音属性；Enter the user's registration information and collect voice samples, specifically by guiding the user to enter preset specific text or answer specific questions to collect the user's voice samples; pre-process the collected voice samples, including denoising, normalization, segmentation and other steps, use voice signal processing technology to extract the features of the voice samples, such as Mel Frequency Cepstral Coefficient (MFCC), pitch, tone, etc., analyze the extracted voice features, and identify the user's unique voice attributes;
根据分析结果，为每个用户生成一个声音模板，该模板包含了用户语音的关键特征，将声音模板和用户特征数据存储在安全的数据库中，与用户账号关联；创建数据索引，以便快速检索和匹配用户的语音特征数据，定期对存储的用户数据进行备份，以防数据丢失或损坏；允许用户定期更新其语音样本，以适应可能的声音变化，定期对存储的语音数据进行维护和清理，删除不再需要的数据或更新过时的信息。Based on the analysis results, a voice template is generated for each user. The template contains the key features of the user's voice. The voice template and user feature data are stored in a secure database and associated with the user account. A data index is created to quickly retrieve and match the user's voice feature data. The stored user data is backed up regularly to prevent data loss or damage. Users are allowed to update their voice samples regularly to adapt to possible voice changes. The stored voice data is regularly maintained and cleaned up, deleting data that is no longer needed or updating outdated information.
语音内容识别：Voice content recognition:
捕获用户的语音，对采集的原始语音信号进行预处理，包括去噪、回声消除、增益控制等，以提高语音质量；从预处理后的语音信号中提取特征，常用的特征包括梅尔频率倒谱系数（MFCC）、音高、音调等，使用训练有声学模型（如隐马尔可夫模型或深度学习模型）来将提取的语音特征映射到对应的声学单元（如音素或字母）；结合语言模型（如n-gram模型或神经网络语言模型）来提高识别的上下文准确性；将声学模型和语言模型的输出进行解码，生成最可能的词序列，根据解码结果，生成一系列候选词或短语；对候选词进行上下文分析，选择最符合当前语境的词或短语，具体过程如下：Capture the user's voice and preprocess the collected original voice signal, including denoising, echo cancellation, gain control, etc., to improve the voice quality; extract features from the preprocessed voice signal. Common features include Mel-frequency cepstral coefficients (MFCC), pitch, tone, etc. Use a trained acoustic model (such as a hidden Markov model or a deep learning model) to map the extracted voice features to corresponding acoustic units (such as phonemes or letters); Combine language models (such as n-gram models or neural network language models) to improve the contextual accuracy of recognition; Decode the outputs of the acoustic model and language model to generate the most likely word sequence, and generate a series of candidate words or phrases based on the decoding results; Perform contextual analysis on the candidate words and select the words or phrases that best fit the current context. The specific process is as follows:
对每个候选词的在当前上下文的语法和语义合适度进行综合分析，通过分析候选词的合适评价参数，其中合适评价参数包括：A comprehensive analysis is performed on the grammatical and semantic suitability of each candidate word in the current context by analyzing the appropriate evaluation parameters of the candidate word, where the appropriate evaluation parameters include:
语法合适度：评估候选词是否符合当前句子的语法结构，将评估结果通过语法合适度评分进行量化，并记为语法值，以此语法值作为衡量语法合适度标准；语义合适度：评估候选词是否在语义上与上下文相匹配，包括词义的一致性和逻辑性，将评估结果通过评分进行量化，并记为语义值，以此语义值作为衡量语义合适度的标准；概率评分：利用语言模型为候选词序列分配概率评分，通过基于n-gram模型或神经网络语言模型；词频统计：考虑候选词在类似上下文中的出现频率，出现频率越高的词更合适；词向量相似度：使用词嵌入技术（如Word2Vec或GloVe）来评估候选词与上下文中其他词的向量空间距离；依赖关系评分：在句法分析的基础上，评估候选词与上下文中其他词的句法依赖关系，并通过赋予依赖关系评分进行量化；Grammatical suitability: Evaluate whether the candidate word conforms to the grammatical structure of the current sentence, quantify the evaluation result through the grammatical suitability score, and record it as a grammatical value, which is used as a standard for measuring grammatical suitability; Semantic suitability: Evaluate whether the candidate word matches the context semantically, including the consistency and logic of the word meaning, quantify the evaluation result through the score, and record it as a semantic value, which is used as a standard for measuring semantic suitability; Probability score: Use the language model to assign probability scores to candidate word sequences, based on the n-gram model or the neural network language model; Word frequency statistics: Consider the frequency of occurrence of candidate words in similar contexts, and the more frequent the word, the more appropriate it is; Word vector similarity: Use word embedding technology (such as Word2Vec or GloVe) to evaluate the vector space distance between the candidate word and other words in the context; Dependency score: Based on syntactic analysis, evaluate the syntactic dependency between the candidate word and other words in the context, and quantify it by assigning a dependency score;
再分别将得到的语法值、语义值、概率评分、出现频率、向量空间距离及依赖关系评分标定为YF、YY、GP、CP、KJ及YP，归一化处理后代入以下公式：以得到合评值HPZ，式中分别为语法值、语义值、概率评分、出现频率、向量空间距离及依赖关系评分的预设权重系数，并以得到的合评值作为衡量目前分析的候选词在当前上下文的语法和语义合适度的综合评价标准；将所有候选词得到的合评值按照大小进行排序，选取合评值最大的候选词作为最优项，同时在用户界面展示最优项的分析结果，并提供选项供用户确认或更正。The obtained grammatical value, semantic value, probability score, occurrence frequency, vector space distance and dependency score are calibrated as YF, YY, GP, CP, KJ and YP respectively, and after normalization, they are entered into the following formula: To obtain the combined evaluation value HPZ, The preset weight coefficients are respectively the grammatical value, semantic value, probability score, occurrence frequency, vector space distance and dependency score, and the obtained combined evaluation value is used as the comprehensive evaluation criterion to measure the grammatical and semantic suitability of the candidate word currently analyzed in the current context; the combined evaluation values obtained for all candidate words are sorted according to size, and the candidate word with the largest combined evaluation value is selected as the optimal option. At the same time, the analysis result of the optimal option is displayed in the user interface, and options are provided for users to confirm or correct.
最后将语音识别结果以文本形式输出，供用户查看或进一步处理。Finally, the speech recognition results are output in text form for users to view or further process.
个性化调整：Personalized Adjustments:
确定识别的语音所属的用户，并调取该用户存储的语音特征，包括口音、语速、常用词汇、语法习惯等，根据收集到的数据，更新用户的个性化模型，包括声学模型和语言模型；对语音内容识别模块输出的结果进行分析，确定可能的识别错误或不确定性，具体过程如下：Determine the user to whom the recognized voice belongs, and retrieve the user's stored voice features, including accent, speaking speed, common vocabulary, grammatical habits, etc. Based on the collected data, update the user's personalized model, including acoustic model and language model; analyze the results output by the voice content recognition module to determine possible recognition errors or uncertainties. The specific process is as follows:
收集语音内容识别模块的输出结果，包括识别的文本和对应的置信度评分，分析每个识别结果的置信度，低于预设置信度表明识别结果存在不确定性或错误；识别常见的错误模式，如常见的替换错误、插入错误或删除错误，检查识别结果是否与上下文信息一致，包括语法、语义和逻辑，通过用户界面让用户能够标记错误或不确定性，并将这些反馈集成到分析过程中；记录所有识别错误和不确定性的实例，用于后续分析和模型训练，分析错误的原因，如声学环境干扰、语言模型限制或声学模型不匹配；将识别结果与候选词列表进行比较，评估替代词或短语，根据错误分析，更新个性化词典和语法规则，以减少未来的错误。Collect the output results of the speech content recognition module, including the recognized text and the corresponding confidence score, analyze the confidence of each recognition result, and if the confidence is lower than the preset confidence, it indicates that there is uncertainty or error in the recognition result; identify common error patterns, such as common substitution errors, insertion errors, or deletion errors, check whether the recognition results are consistent with the context information, including grammar, semantics, and logic, and enable users to mark errors or uncertainties through the user interface, and integrate these feedbacks into the analysis process; record all instances of recognition errors and uncertainties for subsequent analysis and model training, and analyze the causes of the errors, such as acoustic environment interference, language model limitations, or acoustic model mismatch; compare the recognition results with the candidate word list, evaluate alternative words or phrases, and update the personalized dictionary and grammar rules based on the error analysis to reduce future errors.
结合用户的使用上下文和历史数据，增强系统对用户语言习惯的理解；为用户构建个性化词典，包括用户常用的术语、缩写和专有名词，调整识别系统的参数，如敏感度、识别阈值等，以适应用户的语音特征；根据用户的反馈，修正识别错误，并更新系统以避免未来的错误，实施自适应学习算法，使系统能够自动从用户的反馈中学习并改进；再结合用户的其他数据源，如文本输入习惯，以提供更全面的个性化体验，持续监控系统性能，确保个性化优化措施有效并及时调整。Combine the user's usage context and historical data to enhance the system's understanding of the user's language habits; build a personalized dictionary for the user, including the user's commonly used terms, abbreviations and proper nouns, and adjust the recognition system's parameters, such as sensitivity, recognition threshold, etc., to adapt to the user's voice characteristics; based on user feedback, correct recognition errors and update the system to avoid future errors, implement adaptive learning algorithms so that the system can automatically learn and improve from user feedback; combine with other user data sources, such as text input habits, to provide a more comprehensive personalized experience, continuously monitor system performance, ensure that personalized optimization measures are effective and adjusted in a timely manner.
反馈优化：Feedback optimization:
根据提供的用户界面让用户对识别的内容进行反馈，包括识别结果的正确性、建议或错误报告；允许用户对识别结果进行评价，如“正确”或“错误”，并提供更正错误的选项；自动记录系统识别错误，包括用户未确认的识别结果和用户标记的错误，对用户反馈的错误进行人工或自动标注，以便于分析和训练数据的生成；Allow users to provide feedback on the recognized content according to the provided user interface, including the correctness of the recognition results, suggestions or error reports; allow users to evaluate the recognition results, such as "correct" or "wrong", and provide options to correct errors; automatically record system recognition errors, including recognition results that have not been confirmed by the user and errors marked by the user, and manually or automatically annotate errors reported by users to facilitate analysis and the generation of training data;
分析识别错误的原因，如声学环境问题、语言模型不足、用户口音等，使用用户反馈的数据和错误分析结果，更新或重新训练声学模型和语言模型；根据用户的反馈，调整个性化设置，如词典、语法规则等，根据反馈和错误分析，优化识别算法，提高系统的鲁棒性和准确性；利用用户的反馈来学习用户的语音习惯和语言使用模式，实施自适应机制，使系统能够自动根据用户的反馈进行调整；Analyze the causes of recognition errors, such as acoustic environment problems, insufficient language models, user accents, etc., and use user feedback data and error analysis results to update or retrain acoustic models and language models; adjust personalized settings such as dictionaries and grammar rules based on user feedback, and optimize recognition algorithms based on feedback and error analysis to improve the robustness and accuracy of the system; use user feedback to learn users' voice habits and language usage patterns, and implement adaptive mechanisms so that the system can automatically adjust based on user feedback;
管理反馈循环，确保用户反馈被及时处理，并反映在系统改进中；设计激励机制，鼓励用户提供反馈，如积分、奖励等，具体过程如下：Manage the feedback loop to ensure that user feedback is processed in a timely manner and reflected in system improvements; design incentive mechanisms to encourage users to provide feedback, such as points, rewards, etc. The specific process is as follows:
根据用户需求，确定激励机制的类型，如积分、徽章、排行榜、优惠券、免费服务时间等，设计一个积分系统，根据用户反馈的数量和质量给予相应的积分；设定积分兑换奖励的门槛，确保奖励的合理性和系统的可持续性，建立评估机制，确保用户反馈的质量，避免无效或低质量的反馈获得积分；在用户界面中集成激励机制，让用户能够轻松查看积分、奖励和反馈状态，实施系统跟踪用户反馈，自动给予积分，并提供反馈状态更新；设计奖励兑换流程，让用户能够方便地兑换积分获得奖励，向用户展示他们的反馈如何帮助改进系统，增加用户参与感和满足感；建立社区环境，让用户可以看到其他用户的反馈和奖励，增加互动性。Based on user needs, determine the type of incentive mechanism, such as points, badges, rankings, coupons, free service time, etc., design a points system, and give corresponding points according to the quantity and quality of user feedback; set a threshold for redeeming points for rewards to ensure the rationality of rewards and the sustainability of the system, and establish an evaluation mechanism to ensure the quality of user feedback and avoid obtaining points for invalid or low-quality feedback; integrate incentive mechanisms in the user interface so that users can easily view points, rewards, and feedback status, implement a system to track user feedback, automatically give points, and provide feedback status updates; design a reward redemption process so that users can easily redeem points for rewards, show users how their feedback can help improve the system, and increase user participation and satisfaction; establish a community environment so that users can see other users' feedback and rewards to increase interactivity.
确定用户的需求和动机，分析哪些类型的激励措施最有可能促使用户参与反馈，具体的：Identify user needs and motivations and analyze what types of incentives are most likely to get users to participate in feedback, specifically:
通过调研了解用户对不同激励措施的偏好，确定调研方法，可以是在线调查、AI电话访问或焦点论坛；确定调研的用户样本，确保样本具有代表性，在问卷中列出可能的激励措施，如积分、折扣、免费服务、礼品卡等；让用户对激励措施进行评分或排名，以了解他们的偏好，包括开放性问题，让用户描述他们对激励措施的看法和建议；Understand users’ preferences for different incentives through surveys and determine the survey method, which can be online surveys, AI phone interviews, or focus forums; determine the user sample for the survey, ensure that the sample is representative, and list possible incentives in the questionnaire, such as points, discounts, free services, gift cards, etc.; ask users to rate or rank the incentives to understand their preferences, including open-ended questions to let users describe their views and suggestions on incentives;
通过电子邮件、社交媒体、在线平台或纸质问卷等方式分发调研问卷，收集用户的回答，并确保数据的准确性和完整性；对收集到的数据进行整理和分析，找出用户偏好的模式，使用统计方法分析用户反馈，确定最受欢迎的激励措施；识别影响用户偏好的关键因素，如实用性、价值感知等，进行交叉分析，了解不同用户群体（如不同年龄、性别、使用习惯）的偏好差异；将调研结果整理成报告，包括推荐措施类型，根据调研结果，选择一些激励措施进行小规模测试，验证其有效性。Distribute survey questionnaires through email, social media, online platforms or paper questionnaires to collect user responses and ensure the accuracy and completeness of the data; organize and analyze the collected data to find out the patterns of user preferences, use statistical methods to analyze user feedback, and determine the most popular incentives; identify key factors that affect user preferences, such as practicality, perceived value, etc., conduct cross-analysis, and understand the differences in preferences among different user groups (such as different ages, genders, and usage habits); organize the survey results into a report, including the recommended types of measures, and based on the survey results, select some incentives for small-scale testing to verify their effectiveness.
一种语音内容识别系统，包括用户确认模块、用户资料存储模块、语音内容识别模块、个性化优化模块及反馈模块；A voice content recognition system, comprising a user confirmation module, a user data storage module, a voice content recognition module, a personalized optimization module and a feedback module;
用户确认模块用于通过语音内容确认用户；用户资料存储模块用于存储不同用户的声音模板数据；语音内容识别模块用于对用户的语音进行实时的内容识别；个性化优化模块用于根据用户的语音习惯进行内容识别的优化；反馈模块用于根据用户的反馈或系统判断的修正进行不断优化识别内容。The user confirmation module is used to confirm the user through voice content; the user information storage module is used to store the sound template data of different users; the voice content recognition module is used to perform real-time content recognition of the user's voice; the personalized optimization module is used to optimize content recognition according to the user's voice habits; the feedback module is used to continuously optimize the recognition content based on user feedback or corrections made by the system.
以上公开的本发明优选实施例只是用于帮助阐述本发明。优选实施例并没有详尽叙述所有的细节，也不限制该发明仅为的具体实施方式。显然，根据本说明书的内容，可作很多的修改和变化。本说明书选取并具体描述这些实施例，是为了更好地解释本发明的原理和实际应用，从而使所属技术领域技术人员能很好地理解和利用本发明。本发明仅受权利要求书及其全部范围和等效物的限制。The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to only specific implementation methods. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202411195416.3ACN118865951A (en) | 2024-08-29 | 2024-08-29 | A method and system for voice content recognition |
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202411195416.3ACN118865951A (en) | 2024-08-29 | 2024-08-29 | A method and system for voice content recognition |
| Publication Number | Publication Date |
|---|---|
| CN118865951Atrue CN118865951A (en) | 2024-10-29 |
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202411195416.3APendingCN118865951A (en) | 2024-08-29 | 2024-08-29 | A method and system for voice content recognition |
| Country | Link |
|---|---|
| CN (1) | CN118865951A (en) |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN120068036A (en)* | 2025-04-29 | 2025-05-30 | 浪潮智慧供应链科技(山东)有限公司 | Multi-mode digital person identity verification method and system |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104598796A (en)* | 2015-01-30 | 2015-05-06 | 科大讯飞股份有限公司 | Method and system for identifying identity |
| CN104834900A (en)* | 2015-04-15 | 2015-08-12 | 常州飞寻视讯信息科技有限公司 | Method and system for vivo detection in combination with acoustic image signal |
| CN105989264A (en)* | 2015-02-02 | 2016-10-05 | 北京中科奥森数据科技有限公司 | Bioassay method and bioassay system for biological characteristics |
| CN113378134A (en)* | 2021-06-08 | 2021-09-10 | 国科政信科技(北京)股份有限公司 | Identity authentication method and device |
| US20240203428A1 (en)* | 2022-12-16 | 2024-06-20 | Validsoft Limited | Authentication method and system |
| CN118245994A (en)* | 2024-04-10 | 2024-06-25 | 南京龙垣信息科技有限公司 | Multi-mode identity verification system and method |
| CN118553231A (en)* | 2024-07-24 | 2024-08-27 | 南京听说科技有限公司 | Speech recognition method for multiple languages |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104598796A (en)* | 2015-01-30 | 2015-05-06 | 科大讯飞股份有限公司 | Method and system for identifying identity |
| CN105989264A (en)* | 2015-02-02 | 2016-10-05 | 北京中科奥森数据科技有限公司 | Bioassay method and bioassay system for biological characteristics |
| CN104834900A (en)* | 2015-04-15 | 2015-08-12 | 常州飞寻视讯信息科技有限公司 | Method and system for vivo detection in combination with acoustic image signal |
| CN113378134A (en)* | 2021-06-08 | 2021-09-10 | 国科政信科技(北京)股份有限公司 | Identity authentication method and device |
| US20240203428A1 (en)* | 2022-12-16 | 2024-06-20 | Validsoft Limited | Authentication method and system |
| CN118245994A (en)* | 2024-04-10 | 2024-06-25 | 南京龙垣信息科技有限公司 | Multi-mode identity verification system and method |
| CN118553231A (en)* | 2024-07-24 | 2024-08-27 | 南京听说科技有限公司 | Speech recognition method for multiple languages |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN120068036A (en)* | 2025-04-29 | 2025-05-30 | 浪潮智慧供应链科技(山东)有限公司 | Multi-mode digital person identity verification method and system |
| Publication | Publication Date | Title |
|---|---|---|
| Hsu et al. | Speech emotion recognition considering nonverbal vocalization in affective conversations | |
| US9653068B2 (en) | Speech recognizer adapted to reject machine articulations | |
| CN112069484A (en) | Method and system for information collection based on multimodal interaction | |
| CN111986675A (en) | Voice dialogue method, device and computer readable storage medium | |
| CN114360553B (en) | Method for improving voiceprint safety | |
| CN118865951A (en) | A method and system for voice content recognition | |
| CN118173092A (en) | Online customer service platform based on AI voice interaction | |
| Catania et al. | Automatic Speech Recognition: Do Emotions Matter? | |
| CN119357069A (en) | Automatic testing method of intelligent agent language ability based on self-evolution of multilingual model | |
| CN118197299B (en) | Digital human voice recognition method and system based on human-computer interaction | |
| CN119446141A (en) | A conversation interaction method and device based on speech recognition | |
| CN118410147A (en) | Prefix coding and dynamic selection personalized dialogue system based on meta learning | |
| CN118038876A (en) | Speaker confirmation method and equipment based on voice quality self-adaption and class triplet ideas | |
| Jing | [Retracted] Deep Learning‐Based Music Quality Analysis Model | |
| AT&T | ||
| Mirishkar | Towards Building an Automatic Speech Recognition Systems in Indian Context using Deep Learning | |
| CN118296338B (en) | A multimedia terminal teaching interaction method and system | |
| Watanabe et al. | Building speech corpus with diverse voice characteristics for its prompt-based representation | |
| CN119786021B (en) | Multimode multi-factor depression recognition system integrating emotion information | |
| CN120561251B (en) | Voice question-answering method, system, medium and product based on retrieval enhancement generation | |
| Novakovic | Speaker identification in smart environments with multilayer perceptron | |
| Wright | Modelling Prosodic and Dialogue Information for Automatic Speech Recognition | |
| US11798015B1 (en) | Adjusting product surveys based on paralinguistic information | |
| Hassan | A character gram modeling approach towards Bengali Speech to Text with Regional Dialects | |
| Weng | Zero-Shot Face-Based Voice Conversion: Bottleneck-Free Speech Disentanglement With Adversarial Learning in the Real-World Scenario |
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |