Speaker-aware neural network based beamformer for speaker extraction in speech mixtures

Směrovač paprsku založený na neuronové síti poučené o řečníkovi pro extrakci řečníka ze směsi řečových signálů

conference paper

English

Žmolíková Kateřina, Ing., Ph.D. (FIT)
Delcroix Marc (FIT)
Kinoshita Keisuke (FIT)
Higuchi Takuya (FIT)
Ogawa Atsunori (FIT)
Nakatani Tomohiro (FIT)

• http://www.isca-speech.org/archive/Interspeech_2017/pdfs/0667.PDF
• http://www.fit.vutbr.cz/research/groups/speech/publi/2017/zmolikova_interspeech2017_IS170667.pdf PDF

speaker extraction, speaker-aware neural network, beamforming, mask estimation

This article is about the speaker-aware neural network based beamformer for speaker extraction in speech mixtures. In this work, we address the problem of extracting one target speaker from a multichannel mixture of speech. We use a neural network to estimate masks to extract the target speaker and derive beamformer filters using these masks, in a similar way as the recently proposed approach for extraction of speech in presence of noise. To overcome the permutation ambiguity of neural network mask estimation, which arises in presence of multiple speakers, we propose to inform the neural network about the target speaker so that it learns to follow the speaker characteristics through the utterance. We investigate and compare different methods of passing the speaker information to the network such as making one layer of the network dependent on speaker characteristics. Experiments on mixture of two speakers demonstrate that the proposed scheme can track and extract a target speaker for both closed and open speaker set cases.

In this work, we address the problem of extracting one target speaker from a multichannel mixture of speech. We use a neural network to estimate masks to extract the target speaker and derive beamformer filters using these masks, in a similar way as the recently proposed approach for extraction of speech in presence of noise. To overcome the permutation ambiguity of neural network mask estimation, which arises in presence of multiple speakers, we propose to inform the neural network about the target speaker so that it learns to follow the speaker characteristics through the utterance. We investigate and compare different methods of passing the speaker information to the network such as making one layer of the network dependent on speaker characteristics. Experiments on mixture of two speakers demonstrate that the proposed scheme can track and extract a target speaker for both closed and open speaker set cases.

2017

2655–2659

Proceedings of Interspeech, vol. 2017, no. 08, ISSN 1990-9772

Proceedings of Interspeech 2017

International Speech Communication Association

Stocholm

10.21437/Interspeech.2017-667

000457505000551

2-s2.0-85034117887