獨立低階次矩陣分析結合強化學習於頻域音訊分離之分析探討

Abstract

音訊分離一直是訊號處理上想要達成的目標，若能從眾多音訊中擷取自己想要的訊號，在後續也有相當廣泛的應用，本文利用獨立低階次矩陣分析(ILRMA)結合強化學習(Reinforcement Learning)，對未知聲源的訊號進行盲訊號分離。 獨立向量分析透過統計特性對於聲源的獨立性假設進行分離，非負矩陣分解將原本音訊時頻圖分解成音訊時域圖和頻譜圖，同時結合此兩種方法稱為獨立低階次矩陣分析。修正版本的獨立低階次矩陣分析為迭代式的指數參數變化，調整指數參數可有效解決時頻圖特定頻率以上錯置的問題。本研究方法使用修正指數的獨立低階次矩陣分析結合強化學習，透過強化學習的Q-learning動態更動指數參數來達到更好的效果。 本研究針對(1)複數人聲(2)複數樂器(3)歌聲分離進行音訊分離的實驗，實驗將用(1)獨立低階次矩陣分析(2)修正指數的獨立低階次矩陣分析(3)修正指數的獨立低階次矩陣分析結合強化學習這三個方法進行比較，並使用音訊時域圖、音訊時頻圖、計算時間、訊號失真比變化量(∆SI-SDR)來分析分離的好壞程度。 分析結果可以得知修正指數的獨立低階次矩陣分析可以比原始的獨立低階次矩陣分析有明顯的改善，並從音訊時域圖和音訊時頻圖就可以看出差異。然後修正指數的獨立低階次矩陣分析結合強化學習和修正指數的獨立低階次矩陣分析來比較的話，從訊號失真比變化量可以看出加入Q-learning可以讓數值稍微提升，證實本研究可改善音訊分離品質。

Audio source separation has been a goal of signal processing for decades. If we can separate sound clearly, there will be many applications. The research focuses on combining Independent low-rank matrix analysis (ILRMA) and Reinforcement learning (RL) to separate unknown audio signal. Independent vector analysis(IVA) uses independent probability between each audio source to separate audio signal. In the other method, nonnegative matrix factorization(NMF) is based on separating spectrogram to spectral pattern and audio time-varying gain. As a convergence point of these two methods, independent low-rank matrix analysis (ILRMA) has been proposed, which integrates IVA and NMF in a clever way. Then, ILRMA with parametric index number adjustment can solve spectrogram block permutation problem. The research use Q-learning with dynamic change parametric index number to reach better effect. Finally, the research discusses experimental results, such as (1) speech separation (2) musical instrument separation (3) song separation at three different methods (ILRMA, ILRMA with index number adjustment, ILRMA with index number adjustment and Q-learning). Experimental results show ILRMA with index number adjustment and Q-learning can get better audio separation effect.