模型介绍
中文整理简介 Intro
真假声判别模型的设计旨在有效区分音频样本中的真实声音和伪声音,其中判别的类别涵盖男真声、男假声、女真声和女假声四个具体类别。该模型的训练基于计算机视觉(CV)领域的骨干网络,通过将音频数据转换成频谱图并经过微调,以提高网络对不同类别声音的准确识别能力。在训练过程中,采用了包含真实和伪声音样本的数据集,以确保模型能够充分学习并捕捉与男女真假声相关的特征。通过这一方法,模型能够细致地对不同性别和真伪声进行分类,为音频真伪声的准确判别提供了可靠的解决方案。这一模型在语音处理、音乐制作等领域中具有广泛的应用潜力,为音频分析和处理提供了一种高效而精确的工具。其基于计算机视觉原理的训练和微调策略突显了模型在不同领域的适应性和鲁棒性,为进一步研究和应用提供了有益的范例。
The design of the true and false voice discrimination model aims to effectively distinguish between genuine and false voices in audio samples, covering four specific categories: male chest voice, male falsetto voice, female chest voice, and female falsetto voice. The model's training is based on backbone networks from the field of computer vision (CV), by converting audio data into spectrograms and fine-tuning them to enhance the network's ability to accurately recognize different categories of voices. During the training process, a dataset containing both genuine and false voice samples is used to ensure that the model can fully learn and capture features related to true and false voices of both genders. Through this approach, the model can meticulously classify different genders and true and false voices, providing a reliable solution for accurate voice authenticity discrimination. This model has wide-ranging applications in fields such as speech processing and music production, offering an efficient and precise tool for audio analysis and processing. Its training and fine-tuning strategies based on computer vision principles highlight the model's adaptability and robustness across various domains, providing a beneficial paradigm for further research and application.
在线演示 Demo (推理代码)
使用 Usage
:modelscope-code[]{type="sdk"}
维护 Maintenance
:modelscope-code[]{type="git"}
训练结果 Results
Backbone Mel CQT Chroma
:---------------: :----------------------------------: :---------: :---------:
Swin-S V2 ** 0.968 ** 0.268 ** 0.268 **
MaxViT-T 0.820 ** 0.933 ** 0.250
AlexNet ** 0.994 ** ** 0.963 ** ** 0.586 **
ShuffleNet V2 2.0 0.939 0.669 0.222
GoogleNet 0.983 0.274 0.292
MNASNet-A3 0.756 0.260 0.320
SqueezeNet 1.1 0.963 0.900 0.378
Average 0.918 0.610 0.331
最佳结果 Best Result
Loss curve
Training and validation accuracy
Confusion matrix
数据集 Dataset
镜像 Mirror
评估 Evaluation
引用 Cite
优先采用原作者提供的中文模型卡片;仅有英文说明时由闲社AI翻译整理。重要参数、许可证和商用范围请以原文为准。
你将获得什么
完整仓库方式包含该固定版本中的权重、配置、分词器及说明文件;指定版本方式只下载所选文件或整组分片。多模态模型的投影文件、配置等依赖,请按作者说明配齐。
完整仓库可能包含多个权重格式,因此需要的磁盘空间可能大于单一格式的模型大小。下载文件不等于完成模型部署。
使用前须知
运行模型需要的设备、工具与依赖,以原作者说明为准。
是否允许商用、修改和分发,请查看模型许可证。
闲社提供资料与镜像下载方法,不代表已运行模型或完成安全审计。下载的代码应先检查,再执行。
资料或下载有问题?
登录后可提交反馈。
