本项目致力于唱歌辅助系统相关技术的研究，唱声分离、唱声旋律的自动提取、歌词自动转录和歌唱表现的合理评价亟需进一步提高性能。本项目拟研究：（1）基于非负矩阵部分联合分解和基于深度循环神经网络的单声道唱声分离算法；（2）基于唱声谐波增强，结合唱声音高的时域连续性和音高变化的平滑性的自动唱声旋律提取算法；（3）基于特征变换和频谱变换的唱声数据扩展，以深度神经网络-隐马尔可夫模型为基础的唱声建模方案，以优化序列级与识别目标更贴近的目标函数为标准的自动歌词转录算法；（4）基于与主观评价有着较高一致性的特征提取和多特征匹配模型的得分机制，提高歌唱表现的客观评价的合理性。本课题将理论分析和实验研究相结合，旨在建立唱歌辅助系统中唱声分离模型和旋律提取模型，探索有效歌词识别算法，实现唱声的有效分离、旋律的准确提取、歌词转录的准确性和鲁棒性，提高歌唱表现客观评价的合理性，为唱歌辅助系统的应用奠定理论基础。

This project dedicates it to the research of the technology associated with singing assistance. The following four aspect need to be further improved: singing voice separation, singing melody extraction, lyric transcription and the evaluation of singing performance. The project intends to make studies of (1) the monaural singing separation based on Non-negative Matrix Partial Co-Factorization and Deep Recurrent Neural Network; (2) the singing melody extraction based on the enhancement of singing harmonic component and the using of smoothness of pitch varying along the time. (3)the automatic lyrics transcription which includes the singing database extension based on feature transformation and spectrum transformation, and the singing modeling framework based on DNN-HMM (Deep Neural Network-Hidden Markov Model); (4) the evaluation of singing performance that combines the features highly according with the subject evaluation and the scoring mechanism based on the feature matching. This project focuses on the singing separation model and the melody extraction model, combing the theoretical analysis and experimental research, and making study of the lyrics transcription and the rational singing performance evaluation. These researches sever the theoretical foundation of singing assistance system.