Harmful voice detection method in combination with side language information

Abstract

The invention provides a harmful voice detection method in combination with side language information, and the method comprises the steps: collecting a multi-source voice sample, carrying out the preliminary screening and automatic marking of the multi-source voice data through a multi-modal model, and obtaining a preliminary marking sample; performing manual annotation and data reorganization based on the preliminary annotation sample to obtain a high-quality annotation data set, and performing audio unified preprocessing on the high-quality annotation data set to obtain standardized input; based on standardized input, self-supervised high-dimensional features are extracted, and a dual-task model topology is constructed to perform audio recognition; and for the dual-task model topology, carrying out source, category and joint capability training to obtain a joint optimization model so as to output a source label and a type label of harmful information of the to-be-detected audio. According to the method, the optimal performance of a text harmful scene can be considered through high-quality multi-dimensional data support and source and category dual judgment, and the paranguage harmful detection capability is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of speech detection technology, and in particular to a method for detecting harmful speech by incorporating paralinguistic information. Background Technology

[0002] With the rapid development of voice interaction scenarios such as internet social networking, online games, live streaming, and voice customer service, voice information has become an important medium for human communication and information dissemination. However, harmful speech and other issues present in the voice environment have become significant factors threatening the health of the online ecosystem and users' psychological safety. Unlike pure text environments, voice communication not only includes the semantic information carried by text but also a large amount of paralinguistic information, such as tone, emotion, speech rate, timbre, and rhythm. These non-verbal components can carry harmful intentions independently in certain scenarios. For example, offensive information can be conveyed simply through a sarcastic tone, a threatening tone, or a suggestive voice, thereby bypassing existing content moderation systems that rely on text analysis.

[0003] Currently, most online platforms rely on text moderation for malicious information detection, which involves transcribing speech to text and then performing natural language processing. While this method is effective against explicit text attacks, it struggles with malicious information conveyed solely through paralinguistic features within the speech. Although some existing audio malicious detection methods attempt to incorporate acoustic features, they generally remain text-based, with paralinguistic features serving only as supplementary tools. This results in a significant drop in accuracy when the text itself is harmless but the tone of voice is offensive.

[0004] More seriously, existing public datasets have significant shortcomings in detecting harmful paralinguistic content. Most public datasets only consider the harmfulness of the text content during annotation, and do not include a large number of samples that express harmful intent solely through paralinguistics, making it impossible to support the training of detection models with paralinguistic recognition capabilities. Although some studies have collected private datasets containing harmful paralinguistic signals, they have not made them public, and the data construction process lacks transparency, which greatly limits the reproducibility of models and the conduct of comparative studies in the field. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a harmful speech detection method that combines paralinguistic information. This method can achieve optimal performance in harmful text scenarios by using high-quality, multi-dimensional data support and dual determination of source and category, thereby significantly improving the ability to detect harmful paralinguistic speech.

[0006] To achieve the above objectives, the present invention provides the following solution: a method for detecting harmful speech by incorporating paralinguistic information, comprising: Collect multi-source speech samples, and use a multimodal model to perform preliminary screening and automatic annotation of the multi-source speech data to obtain preliminary annotated samples; Based on the initial labeled samples, manual annotation and data compilation are performed to obtain a high-quality labeled dataset. Audio preprocessing is then performed on the high-quality labeled dataset to obtain standardized input. Based on the standardized input, self-supervised high-dimensional features are extracted, and a dual-task model topology is constructed based on the self-supervised high-dimensional features for audio recognition. For the aforementioned dual-task model topology, source identification capability training, category identification capability training, and joint capability training are performed to obtain a joint optimization model, which outputs harmful information source labels and type labels for the audio to be detected.

[0007] Optionally, multi-source speech data is collected, and a multimodal model is used to perform preliminary screening and automatic annotation of the multi-source speech data to obtain preliminary labeled samples, including: Collect real audio samples from multiple publicly available speech corpora, and design prompt words based on the real audio samples to drive a large language model to generate content text that is harmless at the text layer and harmful at the sub-language layer; Based on the content text, speech synthesis technology is used to control roles and styles to obtain synthetic samples including harmful paralinguistic attributes. Then, the synthetic samples and the real audio samples are fused to obtain multi-source speech samples. Using Gemini 1.5 Flash and R-AQA, the multi-source speech samples are judged in three fixed dimensions to obtain the multi-dimensional judgment results of the first model and the multi-dimensional judgment results of the second model. Determine whether the multi-dimensional judgment results of the first model and the multi-dimensional judgment results of the second model are consistent. If they are consistent, perform automatic annotation; otherwise, perform manual annotation to obtain preliminary labeled samples.

[0008] Optionally, the fixed dimensions include a first dimension for determining whether the multi-source speech samples contain harmful content, a second dimension for determining the category of harmful content, and a third dimension for determining the source of harmful content.

[0009] Optionally, based on the initial labeled samples, manual annotation and data compilation are performed to obtain a high-quality labeled dataset. The high-quality labeled dataset then undergoes unified audio preprocessing to obtain standardized input, including: Based on the fixed dimensions and fine-grained categories, the initially labeled samples are manually labeled to obtain secondary labeled samples; Based on the initial labeled sample, a safe sample with harmless text and sub-language is configured at a 1:1 ratio. The safe sample and the secondary labeled sample are combined to obtain a high-quality labeled dataset. Based on the high-quality dataset, resampling, mono conversion, and audio length truncation are performed to unify the audio sample specifications and obtain standardized input.

[0010] Optionally, based on the standardized input, self-supervised high-dimensional features are extracted, and a dual-task model topology is constructed based on the self-supervised high-dimensional features for audio recognition, including: Based on the standardized input, a pre-trained self-supervised speech coding model is used to transform the audio waveform into a self-supervised high-dimensional representation that includes semantic information and paralinguistic information. Based on the self-supervised high-dimensional representation, a multi-label classifier is constructed to predict the source of harmful information, and a multi-class classifier is constructed to predict the category of harmful information or the category of safe information, resulting in a dual-task model topology that includes a source classification head and a category classification head.

[0011] Optionally, for the dual-task model topology, source identification capability training, category identification capability training, and joint capability training are performed to obtain a joint optimization model, which outputs harmful information source labels and type labels for the audio to be detected, including: Based on the dual-task model topology, the category classification head is fixed, the source classification head is trained using the first training set samples, and a multi-label binary cross-entropy loss function is introduced to learn to distinguish the source types of harmful information and complete the source identification capability training. Based on the dual-task model topology, the source classification head is fixed, the category classification head is trained using the second training set samples, and a weighted cross-entropy loss function is introduced to learn the classification of harmful information types and complete the category recognition capability training. Based on the dual-task model topology, using the high-quality labeled dataset, and employing a class-balanced sampling strategy, the source classifier and the class classifier are trained simultaneously to complete joint capability training, resulting in a joint optimization model. The joint optimization model is then used to infer the harmful information source label and type label of the audio to be detected.

[0012] Optionally, the first training set sample is obtained by extracting harmful information, harmful sources, and one-third of the safe samples from the high-quality labeled dataset, and the second training set sample is obtained by extracting text-harmful and sub-language-safe samples, as well as one-seventh of the safe samples from the high-quality labeled dataset.

[0013] This invention discloses the following technical effects by providing a method for detecting harmful speech that incorporates paralinguistic information: 1. Significantly improves the detection capability of harmful paralinguistic information: Compared with existing speech harmful detection methods that mainly rely on text transcription, this invention can maintain a high detection rate in scenarios where the text is harmless but harmful paralinguistic speech exists. On the test subset where only paralinguistic speech is harmful, the detection accuracy is improved by more than 40%, effectively making up for the recognition shortcomings of traditional methods in such scenarios.

[0014] 2. Maintaining optimal performance even in ordinary text-based harmful scenarios: In scenarios where the text itself contains explicit harmful content, the method of this invention still outperforms existing baseline methods, exhibiting higher accuracy and macro-average F1 score. This demonstrates that while introducing para-language detection capabilities, this invention has not weakened its performance on traditional text-based harmful content detection tasks.

[0015] 3. Achieving dual determination of harmful speech source and harmful speech category: Existing harmful speech detection models typically only output the harmful speech category. This invention adds source classification capability, enabling simultaneous determination of whether harmful information originates from text, paralinguistics, or a combination of both. This dual determination capability provides content moderation systems with more refined decision-making basis, such as adopting differentiated processing strategies for content that is harmful only in paralinguistics.

[0016] 4. High-quality, multi-dimensional harmful speech dataset support: The ToxiAlert-Bench dataset constructed in this invention covers 8 publicly available speech corpora and various synthetic samples. The total amount of data, the types of harmful speech sources, and the fine-grained categories are all significantly rich, which can provide reliable and reproducible benchmark support for subsequent paralanguage detection research and practical deployment.

[0017] 5. Dual-head multi-stage training strategy to improve task collaboration: This invention reduces negative transfer between tasks by training the source classification head and the category classification head in stages, improves the independent performance of the two sub-tasks, and then achieves overall optimization through joint fine-tuning, so that source recognition and category recognition can achieve high accuracy at the same time.

[0018] 6. Effectively alleviates class imbalance problem: Introducing class balancing sampling and weighted loss function during training significantly improves the detection performance of the model on small classes and a few source classes, making it more adaptable and more stable and generalizable in real complex scenarios.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the harmful speech detection method provided in an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] like Figure 1 , Figure 2 As shown, the present invention provides a method for detecting harmful speech by incorporating paralinguistic information, comprising: 1. Collect multi-source speech samples, and use a multimodal model to perform preliminary screening and automatic annotation of the multi-source speech data to obtain preliminary labeled samples. Specifically, this includes: 1.1 Collect real audio samples from multiple publicly available speech corpora. Based on the real audio samples, design prompt words to drive the large language model to generate content text that is harmless at the text layer and harmful at the sub-language layer.

[0025] For example, we collect real audio samples containing both harmful and harmless content from eight publicly available speech corpora: TIMIT, IEMOCAP, VCTK, LibriSpeech, VoxCeleb1, LJSpeech, CommonVoice, and MELD. Then, we design prompt words to drive a large language model to generate text that is "harmless at the textual level but harmful at the paralinguistic level."

[0026] 1.2 Based on the content text, speech synthesis technology is used to perform role and style control to obtain a synthesized sample including harmful paralinguistic attributes. The synthesized sample and the real audio sample are then fused to obtain a multi-source speech sample.

[0027] 1.3 Using Gemini 1.5 Flash and R-AQA, the multi-source speech samples are judged in three fixed dimensions to obtain the first model multi-dimensional judgment result and the second model multi-dimensional judgment result; the fixed dimensions include a first dimension for judging whether the multi-source speech samples contain harmful content, a second dimension for judging the category of harmful content, and a third dimension for judging the source of harmful content.

[0028] 1.4 Determine whether the multi-dimensional judgment results of the first model and the multi-dimensional judgment results of the second model are consistent. If they are, perform automatic annotation; otherwise, perform manual annotation to obtain preliminary labeled samples.

[0029] 2. Based on the initial labeled samples, perform manual annotation and data organization to obtain a high-quality labeled dataset. Then, perform unified audio preprocessing on the high-quality labeled dataset to obtain standardized input. Specifically, this includes: 2.1 Based on the fixed dimension and fine-grained category, the preliminary labeled samples are manually labeled to obtain secondary labeled samples.

[0030] Human annotators determine whether each piece of speech is harmful by classifying it into harmful source categories, broad harmful speech categories, and fine-grained categories. The harmful speech source categories are divided into four types: harmful plain text, harmful pure paralanguage, harmful text and paralanguage combined, and safe.

[0031] 2.2 Based on the initial labeled sample, configure a safe sample that is harmless in both text and sub-language at a ratio of 1:1. Combine the safe sample with the secondary labeled sample to obtain a high-quality labeled dataset. The high-quality labeled dataset can be divided into a training set, a validation set, and a test set at a ratio of 7:1:2.

[0032] 2.3 Based on the high-quality dataset, resampling to 16 kHz, mono conversion, and audio length truncation to a maximum of 25 seconds are performed to unify the audio sample specifications and obtain standardized input, which facilitates subsequent feature extraction and model training.

[0033] 3. Based on the standardized input, extract self-supervised high-dimensional features, and construct a dual-task model topology based on the self-supervised high-dimensional features for audio recognition. Specifically, this includes: 3.1 Based on the standardized input, a pre-trained self-supervised speech coding model is used to transform the audio waveform into a self-supervised high-dimensional representation that includes semantic information and paralinguistic information, so as to adapt to the dual tasks of source classification and category classification.

[0034] 3.2 Based on the self-supervised high-dimensional representation, a multi-label classifier is constructed to predict the source of harmful information, and a multi-class classifier is constructed to predict the category of harmful information or the category of safe information, resulting in a dual-task model topology including a source classification head and a category classification head.

[0035] 4. For the aforementioned dual-task model topology, perform source identification capability training, category identification capability training, and joint capability training to obtain a joint optimized model, which outputs the harmful information source label and type label of the audio to be detected. This includes: 4.1 Based on the dual-task model topology, the category classification head is fixed, and the first training set samples are obtained by extracting harmful information, harmful sources and one-third of the safe samples from the high-quality labeled dataset. The source classification head is trained using the first training set samples, and a multi-label binary cross-entropy loss function is introduced to learn to distinguish the source types of harmful information and complete the source identification ability training. 4.2 Based on the dual-task model topology, the source classification head is fixed. By extracting text-harmful and sub-language-safe samples and one-seventh of the safe samples from the high-quality labeled dataset, the second training set samples are obtained. The category classification head is trained using the second training set samples, and a weighted cross-entropy loss function is introduced to learn the classification of harmful information types and complete the category recognition ability training. 4.3 Based on the dual-task model topology, using the high-quality labeled dataset (i.e., the full training set), and employing a class-balanced sampling strategy, the source classification head and the category classification head are trained simultaneously to complete joint capability training, resulting in a joint optimization model. The joint optimization model is then used to infer the harmful information source label and type label of the audio to be detected.

[0036] The class-balanced sampling strategy involves drawing m samples from each class in each batch, with batch size B = m × K (where K is the total number of classes). A preferred embodiment is: m = 3, K = 8, then B = 24.

[0037] Finally, the input speech is detected during the inference stage: the audio to be detected is input into the trained encoder and dual classification head structure, and harmful information source label and harmful category label are output, which can be used for real-time interception, manual review or generation of detailed detection reports.

[0038] Therefore, this invention provides a harmful speech detection method that combines paralinguistic information. It can significantly improve the paralinguistic harmful speech detection capability by taking into account the optimal performance of harmful text scenarios through high-quality multi-dimensional data support and dual determination of source and category.

[0039] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0040] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for detecting harmful speech by incorporating paralinguistic information, characterized in that, include: Collect multi-source speech samples, and use a multimodal model to perform preliminary screening and automatic annotation of the multi-source speech data to obtain preliminary annotated samples; Based on the initial labeled samples, manual annotation and data compilation are performed to obtain a high-quality labeled dataset. Audio preprocessing is then performed on the high-quality labeled dataset to obtain standardized input. Based on the standardized input, self-supervised high-dimensional features are extracted, and a dual-task model topology is constructed based on the self-supervised high-dimensional features for audio recognition. For the aforementioned dual-task model topology, source identification capability training, category identification capability training, and joint capability training are performed to obtain a joint optimization model, which outputs harmful information source labels and type labels for the audio to be detected.

2. The harmful speech detection method combining paralinguistic information according to claim 1, characterized in that, Collect multi-source speech data, and use a multimodal model to perform preliminary screening and automatic annotation of the multi-source speech data to obtain preliminary labeled samples, including: Collect real audio samples from multiple publicly available speech corpora, and design prompt words based on the real audio samples to drive a large language model to generate content text that is harmless at the text layer and harmful at the sub-language layer; Based on the content text, speech synthesis technology is used to control roles and styles to obtain synthetic samples including harmful paralinguistic attributes. Then, the synthetic samples and the real audio samples are fused to obtain multi-source speech samples. Using Gemini 1.5 Flash and R-AQA, the multi-source speech samples are judged in three fixed dimensions to obtain the multi-dimensional judgment results of the first model and the multi-dimensional judgment results of the second model. Determine whether the multi-dimensional judgment results of the first model and the multi-dimensional judgment results of the second model are consistent. If they are consistent, perform automatic annotation; otherwise, perform manual annotation to obtain preliminary labeled samples.

3. The harmful speech detection method combining paralinguistic information according to claim 2, characterized in that, The fixed dimensions include a first dimension for determining whether the multi-source speech samples contain harmful content, a second dimension for determining the category of harmful content, and a third dimension for determining the source of harmful content.

4. The harmful speech detection method combining paralinguistic information according to claim 3, characterized in that, Based on the initial labeled samples, manual annotation and data organization are performed to obtain a high-quality labeled dataset. Audio preprocessing is then applied to this high-quality labeled dataset to obtain standardized input, including: Based on the fixed dimensions and fine-grained categories, the initially labeled samples are manually labeled to obtain secondary labeled samples; Based on the initial labeled sample, a safe sample with harmless text and sub-language is configured at a 1:1 ratio. The safe sample and the secondary labeled sample are combined to obtain a high-quality labeled dataset. Based on the high-quality dataset, resampling, mono conversion, and audio length truncation are performed to unify the audio sample specifications and obtain standardized input.

5. The harmful speech detection method combining paralinguistic information according to claim 4, characterized in that, Based on the standardized input, self-supervised high-dimensional features are extracted, and a dual-task model topology is constructed based on the self-supervised high-dimensional features for audio recognition, including: Based on the standardized input, a pre-trained self-supervised speech coding model is used to transform the audio waveform into a self-supervised high-dimensional representation that includes semantic information and paralinguistic information. Based on the self-supervised high-dimensional representation, a multi-label classifier is constructed to predict the source of harmful information, and a multi-class classifier is constructed to predict the category of harmful information or the category of safe information, resulting in a dual-task model topology that includes a source classification head and a category classification head.

6. The harmful speech detection method combining paralinguistic information according to claim 5, characterized in that, For the aforementioned dual-task model topology, source identification capability training, category identification capability training, and joint capability training are performed to obtain a joint optimization model, which outputs harmful information source labels and type labels for the audio to be detected, including: Based on the dual-task model topology, the category classification head is fixed, the source classification head is trained using the first training set samples, and a multi-label binary cross-entropy loss function is introduced to learn to distinguish the source types of harmful information and complete the source identification capability training. Based on the dual-task model topology, the source classification head is fixed, the category classification head is trained using the second training set samples, and a weighted cross-entropy loss function is introduced to learn the classification of harmful information types and complete the category recognition capability training. Based on the dual-task model topology, using the high-quality labeled dataset, and employing a class-balanced sampling strategy, the source classifier and the class classifier are trained simultaneously to complete joint capability training, resulting in a joint optimization model. The joint optimization model is then used to infer the harmful information source label and type label of the audio to be detected.

7. The harmful speech detection method combining paralinguistic information according to claim 6, characterized in that, The first training set sample is obtained by extracting harmful information, harmful sources, and one-third of the safe samples from the high-quality labeled dataset. The second training set sample is obtained by extracting textually harmful and sub-language-safe samples, as well as one-seventh of the safe samples from the high-quality labeled dataset.

Images