informИнформатикаInformatics1816-03012617-6963UIIP NASB10.37661/1816-0301-2020-17-2-36-43inform-968Research ArticleОБРАБОТКА СИГНАЛОВ, ИЗОБРАЖЕНИЙ, РЕЧИ, ТЕКСТА И РАСПОЗНАВАНИЕ ОБРАЗОВSIGNAL, IMAGE, SPEECH, TEXT PROCESSING AND PATTERN RECOGNITIONВыделение речевой активности на фоне шумов при помощи компактной сверточной нейронной сетиVoice activity detection in noisy conditions using tiny convolutional neural networkВашкевичГ. С.VashkevichR. S.

Вашкевич Григорий Сергеевич, магистр технических наук, аспирант кафедры ЭВС

Минск

Ryhor S. Vashkevich, M. Sci. (Eng.), Postgraduate Student of the Department of EMU

Minsk

ryhorv@gmail.comАзаровИ. С.AzarovE. S.

Азаров Илья Сергеевич, доктор технических наук, доцент, заведующий кафедрой ЭВС

Минск

Elias S. Azarov, Dr. Sci. (Eng.), Associate Professor, Head of the Department of EMU

Minsk

azarov@bsuir.byБелорусский государственный университет информатики и радиоэлектроникиBelarusian State University of Informatics and Radioelectronics2020220420201723643Copyright © Вашкевич Г.С., Азаров И.С., 20202020Вашкевич Г.С., Азаров И.С.Vashkevich R.S., Azarov E.S.Данная работа распространяется под лицензией Creative Commons Attribution 4.0.This work is licensed under a Creative Commons Attribution 4.0 License.https://inf.grid.by/jour/article/view/968

Исследуется задача выделения речевой активности из зашумленного звукового сигнала. Предлагается компактная модель сверточной нейронной сети, которая имеет всего 385 параметров. Модель нетребовательна к вычислительным ресурсам, что позволяет использовать ее в рамках концепции Интернета вещей для портативных устройств с низким энергопотреблением. В то же время эта модель обеспечивает высокую точность определения речевой активности на уровне лучших современных аналогов. Указанные полезные свойства достигаются путем применения специального сверточного слоя, учитывающего гармоническую структуру вокализованной речи и устраняющего избыточность модели за счет инвариантности к изменениям частоты основного тона. В рамках экспериментов производительность модели оценивалась в различных шумовых условиях для разных соотношений сигнала и шума. Результаты экспериментов показали, что предложенная модель обеспечивает более высокую точность определения речевой активности по сравнению с моделью, представленной компанией Google в фреймворке WebRTC.

The paper investigates the problem of voice activity detection from a noisy sound signal. An extremely compact convolutional neural network is proposed. The model has only 385 trainable parameters. Proposed model doesn’t require a lot of computational resources that allows to use it as part of the “internet of things” concept for compact low power devices. At the same time the model provides state of the art results in voice activity detection in terms of detection accuracy. The properties of the model are achieved by using a special convolutional layer that considers the harmonic structure of vocal speech. This layer also eliminates redundancy of the model because it has invariance to changes of fundamental frequency. The model performance is evaluated in various noise conditions with different signal-to-noise ratios. The results show that the proposed model provides higher accuracy compared to voice activity detection model from the WebRTC framework by Google.

детектор речевой активностигармонический сигналсверточная нейронная сетьчастота основного тонаобработка речиvoice activity detectorharmonic signalconvolutional neural networkpitchspeech processingReferencesYoo I.-C., Lim H., Yook D. Formant-based robust voice activity detection. IEEE/ACM Transactions on Audio, Speech and Language Processing (TASLP), 2015, vol. 23, no. 12, рр. 2238–2245. https://doi.org/10.1109/TASLP.2015.2476762Yoo I.-C., Lim H., Yook D. Formant-based robust voice activity detection. IEEE/ACM Transactions on Audio, Speech and Language Processing (TASLP), 2015, vol. 23, no. 12, рр. 2238–2245. https://doi.org/10.1109/TASLP.2015.2476762Pang J. Spectrum energy based voice activity detection. The 7th IEEE Annual Computing and Communication Workshop and Conference (CCWC), Las Vegas, NV, 9–11 January 2017. Las Vegas, 2017, pp. 1–5. https://doi.org/10.1109/CCWC.2017.7868454Pang J. Spectrum energy based voice activity detection. The 7th IEEE Annual Computing and Communication Workshop and Conference (CCWC), Las Vegas, NV, 9–11 January 2017. Las Vegas, 2017, pp. 1–5. https://doi.org/10.1109/CCWC.2017.7868454Kinnunen T., Chernenko E., Tuononen M., Fränti P., Li H. Voice activity detection using MFCC features and support vector machine. The 12th International Conference on Speech and Computer (SPECOM07), Moscow, Russia, 15–18 October 2007. Moscow, 2007, vol. 2, pp. 556–561.Kinnunen T., Chernenko E., Tuononen M., Fränti P., Li H. Voice activity detection using MFCC features and support vector machine. The 12th International Conference on Speech and Computer (SPECOM07), Moscow, Russia, 15–18 October 2007. Moscow, 2007, vol. 2, pp. 556–561.Zazo R., Sainath T. N., Simko G., Parada C. Feature learning with raw-waveform CLDNNs for voice activity detection. 17 th Annual Conference of the International Speech Communication Association, San Francisco, CA, USA, 8–12 September 2016. San Francisco, 2016, pp. 3668–3672. https://doi.org/10.21437/Interspeech.2016-268Zazo R., Sainath T. N., Simko G., Parada C. Feature learning with raw-waveform CLDNNs for voice activity detection. 17 th Annual Conference of the International Speech Communication Association, San Francisco, CA, USA, 8–12 September 2016. San Francisco, 2016, pp. 3668–3672. https://doi.org/10.21437/Interspeech.2016-268Zhang X., Wu J. Denoising deep neural networks based voice activity detection. International Conference on Acoustics, Speech and Signal Processing, Vancouver, BC, Canada, 26–31 May 2013. Vancouver, 2013, pp. 853–857. https://doi.org/10.1109/ICASSP.2013.6637769Zhang X., Wu J. Denoising deep neural networks based voice activity detection. International Conference on Acoustics, Speech and Signal Processing, Vancouver, BC, Canada, 26–31 May 2013. Vancouver, 2013, pp. 853–857. https://doi.org/10.1109/ICASSP.2013.6637769Hughes T., Mierle K. Recurrent neural networks for voice activity detection. International Conference on Acoustics, Speech and Signal Processing, Vancouver, BC, Canada, 26–31 May 2013. Vancouver, 2013, pp. 7378–7382. https://doi.org/10.1109/ICASSP.2013.6639096Hughes T., Mierle K. Recurrent neural networks for voice activity detection. International Conference on Acoustics, Speech and Signal Processing, Vancouver, BC, Canada, 26–31 May 2013. Vancouver, 2013, pp. 7378–7382. https://doi.org/10.1109/ICASSP.2013.6639096Eyben F., Weninger F., Squartini S., Schuller B. Real-life voice activity detection with LSTM Recurrent Neural Networks and an application to Hollywood movies. International Conference on Acoustics, Speech and Signal Processing, Vancouver, BC, Canada, 26–31 May 2013. Vancouver, 2013, pp. 483–487. https://doi.org/10.1109/ICASSP.2013.6637694Eyben F., Weninger F., Squartini S., Schuller B. Real-life voice activity detection with LSTM Recurrent Neural Networks and an application to Hollywood movies. International Conference on Acoustics, Speech and Signal Processing, Vancouver, BC, Canada, 26–31 May 2013. Vancouver, 2013, pp. 483–487. https://doi.org/10.1109/ICASSP.2013.6637694Wang Q., Du J., Bao X., Wang Z.-R., Dai L.-R., Lee C.-H. A universal VAD based on jointly trained deep neural networks. 16 th Annual Conference of the International Speech Communication Association, Dresden, Germany, 6–10 September 2015. Dresden, 2015, рр. 2282–2286.Wang Q., Du J., Bao X., Wang Z.-R., Dai L.-R., Lee C.-H. A universal VAD based on jointly trained deep neural networks. 16 th Annual Conference of the International Speech Communication Association, Dresden, Germany, 6–10 September 2015. Dresden, 2015, рр. 2282–2286.Ryant N., Liberman M., Yuan J. Speech activity detection on youtube using deep neural networks. 14 th Annual Conference of the International Speech Communication Association, Lyon, France, 25–29 August 2013. Lyon, 2013, pp. 728–731.Ryant N., Liberman M., Yuan J. Speech activity detection on youtube using deep neural networks. 14 th Annual Conference of the International Speech Communication Association, Lyon, France, 25–29 August 2013. Lyon, 2013, pp. 728–731.Snyder D., Chen G., Povey D. Musan: a Music, Speech, and Noise Corpus, 2015. Available at: https://arxiv.org/abs/1510.08484 (accessed 20.10.2019).Snyder D., Chen G., Povey D. Musan: a Music, Speech, and Noise Corpus, 2015. Available at: https://arxiv.org/abs/1510.08484 (accessed 20.10.2019).Kasi K., Zahorian S. A. Yet another algorithm for pitch tracking. International Conference on Acoustics, Speech, and Signal Processing, Orlando, 13–17 May 2002. Orlando, 2002, vol. 1, рр. 361–364. https://doi.org/10.1109/ICASSP.2002.5743729Kasi K., Zahorian S. A. Yet another algorithm for pitch tracking. International Conference on Acoustics, Speech, and Signal Processing, Orlando, 13–17 May 2002. Orlando, 2002, vol. 1, рр. 361–364. https://doi.org/10.1109/ICASSP.2002.5743729Kingma D. P., Ba J. Adam: a Method for Stochastic Optimization, 2014. Available at: https://arxiv.org/abs/1412.6980 (accessed 20.10.2019).Kingma D. P., Ba J. Adam: a Method for Stochastic Optimization, 2014. Available at: https://arxiv.org/abs/1412.6980 (accessed 20.10.2019).

The authors declare that there are no conflicts of interest present.