Jinyu Han

Audio Imputation

By Jinyu Han, Gautham J. Mysore and Bryan Pardo

Work presented at LVA/ICA 2012

Tel-Aviv, Israel

March 12-15, 2012.

We present an approach that estimates missing values in the time-frequency domain of audio signals.

  • The proposed approach, based on the Non-negative Hidden Markov Model as explained in detail [here], enables more temporally-coherent estimation for the missing data by accounting for both the spectral and temporal information of the audio signal.

  • This approach is able to reconstruct highly corrupted audio with large parts of the spectrogram missing, and has advantages over performing imputation using PLCA.
  • Related paper

    J. Han, G. J. Mysore, and B. Pardo: Audio Imputation using the Non-negative Hidden Markov Model. Lecture Notes in Computer Science: Latent Variable Analysis and Signal Separation (LVA/ICA), 2012. (Spotlight Presentation) [pdf] [BibTex]

    Modeling of Audio

    Non-negative Spectrogram Factorization (PLCA)

    Non-negative spectrogram factorization refers to a class of methods including non-negative matrix factorization and probabilistic latent component analysis (PLCA), which are used to factorize spectrograms. In this discussion, we will use the specific case of PLCA. However, the ideas generalize to most such methods.

    One of the problems with PLCA is that:

  • It models the audio with a single non-negative dictionary.
  • It does not account for non-stationarity and temporal dynamics of audio.

  • A Large dictionary learned by PLCA

    Non-negative Hidden Markov Model (N-HMM)

  • Model the audio with multiple dictionaries such that each time frame is a linear combination of elements from any one dictionary.
  • A Markov chain models the transitions between dictionaries.

  • Multiple dictionaries and a Markov chain learned by N-HMM.

    An illustration of the proposed audio imputation system using N-HMM (of five states/dictionaries)

    Illustration Examples

    In the following examples we automatically fill in the time-frequency domain using our proposed method and compare the results to a method based Probabilistic Latent Component Analysis. Here are two examples with large regions of the spectrogram missing.

    "Scar Tissue" by "Red Hot Chili Peppers"

    In this example, PLCA produces a reconstruction with a lots of high frequency noise, while the reconstruction by the proposed method is much more clean. Although the reconstructed signal by the proposed method sounds less full in the high frequency range, we still find that it is more perceptually pleasing than adding extra noise in the high frequency domain

    "Born to be wild" by "Steppenwolf"

    In this example, PLCA gets the temporal dynamics of the audio wrong. It is obviously to hear that too much energy is asigned to the percussion sound (the bright vertical strip in the spectrogram) in the music. In contrast, the reconstructed signal by our method has a better temporal dynamics.

    More examples

    "Wonderwall" by "Oasis"

    "She will be loved" by "Maroon 5"

    "Bad Day" by "Daniel Powter"

    "Here I go again" by "Whitesnake"

    "Better together" by "Jack Johnson"

    "1979" by "Smashing Pumpkins"

    "Every Breath You Take" by "Police"


    This work was supported in part by NSF grant numbers IIS-0643752.

    image image image

    This project is a collaboration between Interactive Audio Lab of Northwestern University and Adobe Advanced Technology Labs.