SOUND SOURCE SEPARATION LEARNING DEVICE, SOUND SOURCE SEPARATION LEARNING METHOD, AND SOUND SOURCE SEPARATION LEARNING PROGRAM

• Main Authors: RI RI, WATANABE CHIHIRO, KAMEOKA HIROKAZU, SEKI SHOGO
• Format: Patent
• Language: English; Japanese
• Published: 30.03.2023
• Subjects: ACOUSTICS; CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; MUSICAL INSTRUMENTS; PHYSICS; SPEECH ANALYSIS OR SYNTHESIS; SPEECH OR AUDIO CODING OR DECODING; SPEECH OR VOICE PROCESSING; SPEECH RECOGNITION

To improve the learning efficiency while preventing the sound source separation accuracy from degrading even in a case of separating a mixed signal in which there are many sound sources mixed in learning of a sound source separation model.SOLUTION: A separation unit 101 separates a mixed signal in which a plurality of sound source signals are mixed into a plurality of separation signals by a sound source separation model 111 formed by a neural network. A calculation unit 102 extracts a feature quantity by an extractor 112 from each of the plurality of separation signals and each of the plurality of sound source signals, and calculates an attention matrix having, as an element, similarity between each of the plurality of separation signals and each of the plurality of sound source signals based on the extracted feature quantity. A learning unit 103 learns parameters of the sound source separation model and the extractor so as to minimize a learning standard including an error between each of the plurality of sound source signals and each of the plurality of separation signals associated based on the attention matrix.SELECTED DRAWING: Figure 3 【課題】音源分離モデルの学習において、混合されている音源の数が多い混合信号を分離する場合でも、音源分離精度の低下を抑制しつつ、学習効率を向上させる。【解決手段】分離部１０１が、複数の音源信号が混合された混合信号を、ニューラルネットワークで構成された音源分離モデル１１１により複数の分離信号に分離し、算出部１０２が、複数の分離信号の各々及び複数の音源信号の各々から抽出器１１２により特徴量を抽出し、抽出した特徴量に基づいて、複数の分離信号の各々と複数の音源信号の各々との類似度を要素に持つ注意行列を算出し、学習部１０３が、注意行列に基づいて対応付けした複数の音源信号の各々と複数の分離信号の各々との誤差を含む学習規準を最小化するように、音源分離モデル及び抽出器のパラメータを学習する。【選択図】図３