Modeling speech with sum-product networks: Application to bandwidth extension

Peharz, Robert; Kapeller, Georg; Mowlaee, Pejman; Pernkopf, Franz

doi:10.1109/icassp.2014.6854292

Cited by 44 publications

(36 citation statements)

References 16 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…For re-synthesizing time signals from the log-magnitude spectrogram reconstructions we use the same approach as in [25], using 100 iterations of the Griffin&Lim algorithm [76] to synthesize phase for the reconstructed time-frequency bins. We replace the predicted telephone band by the original telephone band in all models and initialize Griffin&Lim with the corresponding phase.…”

Section: A Setupmentioning

confidence: 99%

“…For objective evaluation, we use the log-spectral distortion (LSD) in the high-band, similar as in [55], [25]. Using 9 order LPC analysis of each frame we get the spectral envelope as (2) where is the square-root of the variance of the LPC-analyzed signal and are the LPC coefficients.…”

Section: B Evaluation Objectivesmentioning

confidence: 99%

“…deep belief networks [19]. Recently, two particularly interesting representation models have been introduced, namely generative stochastic networks (GSNs) [9], [20] and sum-product networks (SPNs) [21], [22], [23], [24], [25], [26].…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Representation Learning for Single-Channel Source Separation and Bandwidth Extension

Zöhrer

Peharz

Pernkopf

2015

IEEE/ACM Trans. Audio Speech Lang. Process.

Self Cite

View full text Add to dashboard Cite

In this paper, we use deep representation learning for model-based single-channel source separation (SCSS) and artificial bandwidth extension (ABE). Both tasks are ill-posed and source-specific prior knowledge is required. In addition to well-known generative models such as restricted Boltzmann machines and higher order contractive autoencoders two recently introduced deep models, namely generative stochastic networks (GSNs) and sum-product networks (SPNs), are used for learning spectrogram representations. For SCSS we evaluate the deep architectures on data of the 2 CHiME speech separation challenge and provide results for a speaker dependent, a speaker independent, a matched noise condition and an unmatched noise condition task. GSNs obtain the best PESQ and overall perceptual score on average in all four tasks. Similarly, frame-wise GSNs are able to reconstruct the missing frequency bands in ABE best, measured in frequency-domain segmental SNR. They outperform SPNs embedded in hidden Markov models and the other representation models significantly. Index Terms-Bandwidth extension, deep neural networks (DNNs), generative stochastic networks, representation learning, single-channel source separation (SCSS), sum-product networks.Manuscript

show abstract

Section: A Setupmentioning

confidence: 99%

Section: B Evaluation Objectivesmentioning

confidence: 99%

See 1 more Smart Citation

Representation Learning for Single-Channel Source Separation and Bandwidth Extension

Zöhrer

Peharz

Pernkopf

2015

IEEE/ACM Trans. Audio Speech Lang. Process.

Self Cite

View full text Add to dashboard Cite

show abstract

“…ABE approaches based on source-filter modelling estimate separate spectral envelope and excitation components [2,3]. Other approaches operate directly on complex short-term spectral estimates derived, e.g., using the Fourier transform (STFT) [4,5] or constant-Q transform [6]. Complementary to short-term spectral estimates, is some form of contextual information, or memory which can be harnessed to improve the reliability of HB component estimation.…”

Section: Introductionmentioning

confidence: 99%

Artificial Bandwidth Extension with Memory Inclusion Using Semi-supervised Stacked Auto-encoders

2018

View full text Add to dashboard Cite

Artificial bandwidth extension (ABE) algorithms have been developed to improve quality when wideband devices receive speech signals from narrowband devices or infrastructure. The utilisation of contextual information in the form of dynamic features or explicit memory captured from neighbouring frames is common to ABE research, however the use of additional cues augments complexity and can introduce latency. Previous work shows that unsupervised, linear dimensionality reduction techniques help to reduce complexity. This paper reports a semisupervised, non-linear approach to dimensionality reduction using a stacked auto-encoder. In further contrast to previous work, it operates on raw spectra from which a low dimensional narrowband representation is learned in a data-driven manner. Three different objective speech quality measures show that the new features can be used with a standard regression model to improve ABE performance. Improvements in the mutual information between learned features and missing higher frequency components are also observed whereas improvements in speech quality are corroborated by informal listening tests.

show abstract

“…The SPN is regarded as observation models in HMMs modeling. The algorithm is robust and computationally inexpensive [10]. Deep learning has emerged as a new area of machine learning research [11].…”

Section: Introductionmentioning

confidence: 99%

A Novel Research to Artificial Bandwidth Extension Based on Deep BLSTM Recurrent Neural Networks and Exemplar-Based Sparse Representation

Liu

Tao

2016

Interspeech 2016

View full text Add to dashboard Cite

This paper presents a two stages artificial bandwidth extension (ABE) framework which combine deep bidirectional Long Short Term Memory (BLSTM) recurrent neural network with exemplar-based sparse representation to estimate missing frequency band. It demonstrates the suitability of proposed method for modeling log power spectra of speech signals in ABE. The BLSTM-RNN which can capture information from anywhere in the feature sequence is used to estimate the log power spectra in the high-band firstly and the exemplar-based sparse representation which could alleviate the oversmoothing problem is applied to generated log power spectra in the second stage. In addition, rich acoustic features in the low-band are considered to reduce the reconstruction error. Experimental results demonstrate that the proposed framework can achieve significant improvements in both objective and subjective measures over the different baseline methods.

show abstract

Modeling speech with sum-product networks: Application to bandwidth extension

Cited by 44 publications

References 16 publications

Representation Learning for Single-Channel Source Separation and Bandwidth Extension

Representation Learning for Single-Channel Source Separation and Bandwidth Extension

Artificial Bandwidth Extension with Memory Inclusion Using Semi-supervised Stacked Auto-encoders

A Novel Research to Artificial Bandwidth Extension Based on Deep BLSTM Recurrent Neural Networks and Exemplar-Based Sparse Representation

Contact Info

Product

Resources

About