Single-channel speech separation using sequential discriminative dictionary learning

Xu, Yangfei; Bao, Guangzhao; Xu, Xu; Ye, Zhongfu

doi:10.1016/j.sigpro.2014.07.012

Cited by 19 publications

(11 citation statements)

References 18 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…A discriminative approach for NMF is proposed in [7] which attempts to minimize the cross-coherence between the basis vectors pertaining to different sources and in [10] where the test-time objective for separation is incorporated in the formulation. The work in [11] and [12] also propose discriminative model based method where the models for the underlying sources are learnt in the form of overcomplete dictionaries. The proposed work in [11] attempts to learn the dictionaries for all the sources simultaneously rather than as independent units and [12] learns a sequence of dictionaries and performs separation in few stages.…”

Section: Related Workmentioning

confidence: 99%

“…For this paper, we have worked with NMF dictionaries as in [7] and have improved their performance by determining better choices for the number of basis vectors and by eventually separating one source at a time. We show through simulations that our framework on the NMF based source separation in [7] outperforms other approaches like [11] and [12] for both speech-speech and speech-music separation tasks.…”

Section: Motivation and Contributionmentioning

confidence: 99%

“…Results show that the quality of separation can be improved with the framework. We also compare our results with the dictionary learning based approach in [12], called SDDL, which extracts the sources in a number of levels to achieve better separation. We compared the performances of the algorithms using SDR and SIR as the evaluation metrics.…”

Section: Performance Of Our Algorithmmentioning

confidence: 99%

See 2 more Smart Citations

On Discriminative Framework for Single Channel Audio Source Separation

Gang

Biyani

2016

Interspeech 2016

View full text Add to dashboard Cite

Single channel source separation (SCSS) algorithms that utilise discriminative source models perform better in comparison to those that are trained independently. However, all the aspects of training discriminative models have not been addressed in the literature. For instance, the choice of dimensions of source models (number of columns of NMF, Dictionary etc) not only influences the fidelity of a given source but also impacts the interference introduced in it. Therefore choosing a right dimension parameter for every source model is crucial for an effective separation. In fact, the similarity between the constituent sources can be different for different mixtures and thus, dimensions should also be chosen specific to the sources in the concerned mixture. Further, separation of a given constituent from a mixture, assuming remaining to be interferers, offers more freedom for the particular constituent and hence provide better separation. In this paper, we propose a generic discriminative learning framework where we separate one source at a time and embed our dimension search algorithm in the training of discriminative source models. We apply our framework on the NMF based SCSS algorithms and demonstrate a performance improvement in separation for both speech-speech and speechmusic mixture.

show abstract

Section: Related Workmentioning

confidence: 99%

Section: Motivation and Contributionmentioning

confidence: 99%

Section: Performance Of Our Algorithmmentioning

confidence: 99%

See 1 more Smart Citation

On Discriminative Framework for Single Channel Audio Source Separation

Gang

Biyani

2016

Interspeech 2016

View full text Add to dashboard Cite

show abstract

“…Otherwise, the sparse representation will not occur. Signal representation using sparse dictionaries are considered in several recent researches as an efficient method for different audio processing issues such as audio structure analysis, automatic music transcription, and audio source separation [4][5][6][7][8][9].…”

Section: Introductionmentioning

confidence: 99%

Representation of Musical Signals using Instrument-Specific Dictionaries

Azamian¹

2017

IJCA

View full text Add to dashboard Cite

A new simple method is proposed to synthesize the instrument-specific dictionaries and its use is examined in the time domain musical signal representation. By investigating the spectrum of musical note signals, it is seen that only a small number of frequency elements are significant in the inherent structure of a musical note, and other elements could be omitted. This sparsity is utilized to synthesize note-specific atoms. Firstly, some basic functions are defined from the long-term spectrum of the note signal, called primary atoms. Then the primary atoms that satisfy some conditions are selected as basic atoms and are incorporated to synthesize note-specific atoms. Some usual signal processing windows also are examined such as Gaussian and Hamming windows to synthesize note-specific atoms. The note-specific atoms of an instrument are integrated in an instrument-specific dictionary. A musical signal is represented by mapping to this dictionary by means of the Matching Pursuit algorithm. The proposed method was evaluated on the RWC musical sound database. The results showed that it improves the quality of signal representation compared to some previous methods.

show abstract

“…Dictionary learning for sparse representation of signals is one of the active topics in various areas such as compression, denoising, inpainting, super-resolution, classification, and source separation [1][2][3][4][5]. Dictionary is a collection of atoms which can represent each training data by a linear combination of the atoms.…”

Section: Introductionmentioning

confidence: 99%