Lock in $30 Savings on PRO—Offer Ends Soon! ⏳

空間音響処理における物理法則に基づく機械学習

 空間音響処理における物理法則に基づく機械学習

人工知能学会 AIチャレンジ研究会 招待講演

Avatar for NII S. Koyama's Lab

NII S. Koyama's Lab

December 01, 2025
Tweet

More Decks by NII S. Koyama's Lab

Other Decks in Research

Transcript

  1. 空間音響処理 ➢ 空間音響収録・再生 – Sound field is captured by multiple

    mics and reproduced by headphones or loudspeakers – Head-related transfer function personalization ➢ 室内音響解析・制御 – Visualization and auralization of spatial sound – Estimation of room acoustic impulse responses/transfer functions ➢ 音源強調・分離 – Beamforming techniques require accurate steering vectors (array manifold vectors) – Source enhancement using wearable devices is more challenging December 2, 2025 4 複数のマイクを用いた音空間の解析・制御
  2. 応用1: マイクアレイ信号からのバイノーラル再現 December 2, 2025 6 VR音響のためのマイクアレイによる空間音響収録とそのバイノーラル再現 Recording Reproduction ➢

    Unlike binaural synthesis in VR space, binaural reproduction in real environments requires spatial audio capturing by using multiple mics ➢ Required to estimate spatial sound in a wide area to achieve a wide listening area, e.g., 6DoF reproduction
  3. 応用2: ステアリングベクトルのアップサンプリング ➢ Estimation of steering vectors for wearable devices

    with multiple mics is crucial for source enhancement compared to simple-shaped arrays ➢ Upsampling techniques for steering vectors will simplify the measurement of steering vectors December 2, 2025 7 インパルス応答測定によるステアリングベクトルの空間的補間
  4. 応用3: 空間アクティブ騒音制御 ➢ Active noise control (ANC) aims to cancel

    noise by using loudspeaker signals, but its effect is limited to local region ➢ Spatial ANC by estimating spatial sound using multiple mics and synthesizing anti- spatial sound using multiple loudspeakers December 2, 2025 8 スピーカ信号による3次元領域内の騒音抑制 Quiet zone
  5. 音場推定 December 2, 2025 9 音場推定の内部/外部問題 Microphone Target region: Target

    region: Microphone 内部問題 外部問題 ここでは内部問題に焦点を置く
  6. 音場推定 December 2, 2025 10 音場推定問題の定式化 Estimate pressure distribution in

    the time domain or in frequency domain with ominidirectional mics at Microphone Target region:
  7. 音場推定 ➢ 一般的な関数補間としての問題設定 – is represented by model parameters December

    2, 2025 11 Formulation of sound field estimation problem Loss term Regularization term 音場推定問題の定式化 Microphone Target region: Observation Samples in space/time/freq
  8. 音場推定 ➢ 一般的な関数補間としての問題設定 – is represented by model parameters December

    2, 2025 12 Formulation of sound field estimation problem Squared ℓ 2 -norm penalty 音場推定問題の定式化 Microphone Target region: Squared error loss
  9. ニューラルネットワーク以前の音場推定法 ➢ 物理的な制約を陽に用いた手法が提案されてきた歴史がある – 基底関数展開に基づく方法 [Williams+ 1999, Colton+ 2013] •

    Plane wave expansion (or Herglotz wave function) • Spherical wave function expansion • Equivalent source distribution (or single-layer potential) – 無限次元展開あるいはカーネル回帰に基づく方法 • Harmonic analysis of infinite order [Ueno+ 2018] • Directionally-weighted kernel regression [Ueno+ 2021] December 2, 2025 16 従来の音場推定に関する包括的なレビュー論文: • Ueno and Koyama, “Sound Field Estimation: Theories and Applications,” Foundations and Trends®️ in Signal Processing, 2025.
  10. 支配方程式の要素解に対する基底関数展開 ➢ Function is modeled by basis functions and their

    weights ➢ 波動方程式/ヘルムホルツ方程式の要素解による基底関数 [Williams+ 1999, Colton+ 2013] – Plane wave expansion (Herglotz wave function) – Spherical wave function expansion – Equivalent source distribution (single-layer potential) December 2, 2025 17 有限個の基底関数の線形結合による表現
  11. 支配方程式の要素解に対する基底関数展開 ➢ Spherical Bessel function December 2, 2025 20 2

    4 6 8 10 12 14 x -0.2 0 0.2 0.4 0.6 0.8 1 n=0 n=2 n=4 n=6 Bessel function
  12. 支配方程式の要素解に対する基底関数展開 ➢ 有限次元の基底関数を用いた線形回帰 – Regularized least squares solution of expansion

    coefs – Estimate the function December 2, 2025 23 基底関数の数や展開中心を適切に設定することが必要
  13. 支配方程式の制約を用いたカーネル回帰 ➢ is represented by weighted sum of kernel function

    ➢ Kernel function is a similarity function expressed as innter product on some functional space December 2, 2025 24 を無限次元とすることや を直接設計することも可能
  14. 支配方程式の制約を用いたカーネル回帰 ➢ In kernel ridge regression, is obtained as with

    Gram matrix defined as ➢ Estimate the function December 2, 2025 25 関数空間 とカーネル関数 を適切に定義することが必要
  15. 支配方程式の制約を用いたカーネル回帰 ➢ Inner product and norm over are defined by

    plane wave expansion with positive directional weighting [Ueno+ 2021] December 2, 2025 26 推定解をヘルムホルツ方程式の解空間に制約するためのカーネル関数 指向性重み関数 は音場の指向特性に 関する事前情報を組み入れるように設計
  16. 支配方程式の制約を用いたカーネル回帰 ➢ Kernel function when is defined by using von

    Mises–Fisher distribution ➢ When no prior information, i.e., uniform weight , December 2, 2025 27 with 推定解をヘルムホルツ方程式の解空間に制約するためのカーネル関数
  17. 支配方程式の制約を用いたカーネル回帰 ➢ Experimental results using real data from MeshRIR dataset

    – Reconstructing pulse signal from single loudspeaker w/ 18 mic December 2, 2025 28 Ground truth Kernel regression w/ HE constraint Kernel regression w/ Gaussian kernel (Black dots indicate mic positions) [Koyama+ 2021]
  18. ニューラルネットワークを用いた音場推定 ➢ High representational power – Solution space in basis

    expansion and kernel regression is highly constrained – High adaptability to the target acoustic environment can be expected by using NNs ➢ From snapshot-based to learning-based – Basically, linear and kernel regressions use only a snapshot observation – Properties of the target acoustic environment can be learned from training data December 2, 2025 29 なぜ音場推定においてニューラルネットワークか? マイク数が極めて少数の場合などに高い推定精度を実現することが期待できる
  19. 支配方程式制約を組み入れた回帰のためのNN ➢ Regression by NNs – Target output is discretized

    as – NN with input and output is designed with NN params – NN is trained using a pair of datasets to minimize the loss, e.g., December 2, 2025 30
  20. 支配方程式制約を組み入れた回帰のためのNN ➢ NNによる基底関数の展開係数の推定 – Train a NN estimating weights of

    basis expansion – Continuous function can be reconstructed by using estimated expansion coefs – Can be regarded as physics-constrained neural network (PCNN) [Karakonstantis+ 2023, Lobato+ 2024] ➢ (近似的な)PDE lossの導入 – Loss function evaluating deviation from governing PDEs: PDE loss – Because of discrete output values, PDE loss is computed by finite difference or interpolation – In [Shigemi+ 2022], physics-informed convolutional neural network (PICNN) using bi- cubic spline interpolation is proposed December 2, 2025 31 どのようにNNに支配方程式の制約を組み入れるか?
  21. 陰的表現のためのNNを用いたPINN ➢ Implicit neural representation [Sitzmann+ 2020] – NNs are

    used to implicity represent a continuous function – NN with input and output is designed with NN params – NN is trained for approximaging by using training data December 2, 2025 32
  22. 陰的表現のためのNNを用いたPINN ➢ Physics-informed neural network (PINN) [Raissi+ 2019] – Implicit

    neural representation allows incorporating constraints on including its (partial) derivatives in loss function December 2, 2025 33 自動微分を用いて計算可能
  23. 陰的表現のためのNNを用いたPINN ➢ Physics-informed neural network (PINN) [Raissi+ 2019] – Case

    when estimating function approximately satisfying Helmhotz eq December 2, 2025 34 Helmholtz方程式からの逸脱度を 評価する損失関数
  24. 現在のPIMLに基づく音場推定手法 December 2, 2025 36 Snapshot-based Learning-based Constrained Penalized [Ribeiro+

    2024] [Karakonstantis+ 2023] [Olivieri+ 2024] [Shigemi+ 2022] PI-strategy Training-strategy [Labato+ 2024] [Chen+ 2023] [Ma+ 2024] [Karakonstantis+ 2024] [Masuyama+ 2025]
  25. Physics-Constrained Neural Kernel ➢ Directional weighting function of kernel function

    is adapted to environment December 2, 2025 37 陰的表現のためのNNによるHelmholtz方程式制約下でのカーネル関数 Directed component Residual component Kernel function based on plane wave expansion [Ribeiro+ 2024]
  26. Physics-Constrained Neural Kernel ➢ Directed component – Weighted sum of

    (sparse) von Mises–Fisher distributions to represent direct sound and early reflections December 2, 2025 38 Sparsity constraint Normalization const 陰的表現のためのNNによるHelmholtz方程式制約下でのカーネル関数
  27. Physics-Constrained Neural Kernel ➢ Residual component – Implicit neural representation

    to represent late reverberation December 2, 2025 39 Computed by numerical integration : Implicit neural representation 陰的表現のためのNNによるHelmholtz方程式制約下でのカーネル関数
  28. Physics-Constrained Neural Kernel ➢ Kernel function is sum of directed

    and residual kernels – Hyperparameters are jointly optimized by a steepest descent-based algorithm – Solution still satisfies Helmholtz equation – Inference by linear operation based on kernel ridge regression December 2, 2025 40 Directed kernel Residual kernel 推定は時間領域でのFIRフィルタとして実現可能 陰的表現のためのNNによるHelmholtz方程式制約下でのカーネル関数
  29. Physics-Constrained Neural Kernel ➢ Numerical experiment: T60: 400 ms, #

    mics: 41, spherical shell array December 2, 2025 41 [Koyama+ 2025] Proposed PCNK Proposed PCNK
  30. まとめ ➢ 空間音響処理におけるPIML – 空間音響処理の基盤となる音場推定におけるPIMLについて解説 – 関数補間に対して物理的な性質を組み入れるアプローチ • 支配方程式の要素解に対する基底関数展開 •

    支配方程式の制約を用いたカーネル回帰 • 支配方程式制約を組み入れた回帰のためのNN • 陰的表現のためのNNを用いたPINN – 現在のPIMLに基づく音場推定 • Physics-Constrained Neural Kernel December 2, 2025 42 Thank you for your attention!