Serban Mihalache

Mr. Serban Mihalache

Speech and Dialogue Research Laboratory (SpeeD), University Politehnica of Bucharest (UPB)
+1
orcid

Research Focus & Metrics

Academic Career Length

12Years (2015-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
6 February 2022 in Sensors
Verified
Journal Article
Open access
PDF file
Serban MihalacheSerban Mihalache
Serban Mihalache
Dragos BurileanuDragos Burileanu
Dragos Burileanu
In this work, we first propose a deep neural network (DNN) system for the automatic detection of speech in audio signals, otherwise known as voice activity detection (VAD). Several DNN types were investigated, including multilayer perceptrons (MLPs), recurrent neural networks (RNNs), and convolutional neural networks (CNNs), with the best performance being obtained for the latter. Additional postprocessing techniques, i.e., hysteretic thresholding, minimum duration filtering, and bilateral extension, were employed in order to boost performance. The systems were trained and tested using several data subsets of the CENSREC-1-C database, with different simulated ambient noise conditions, and additional testing was performed on a different CENSREC-1-C data subset containing actual ambient noise, as well as on a subset of the TIMIT database. An accuracy of up to 99.13% was obtained for the CENSREC-1-C datasets, and 97.60% for the TIMIT dataset. We proceed to show how the final VAD system can be adapted and employed within an utterance-level deceptive speech detection (DSD) processing pipeline. The best DSD performance is achieved by a novel hybrid CNN-MLP network leveraging a fusion of algorithmically and automatically extracted speech features, and reaches an unweighted accuracy (UA) of 63.7% on the RLDD database, and 62.4% on the RODeCAR database.

13 October 2021
Verified
Conference Paper
Closed access
Serban MihalacheSerban Mihalache
Serban Mihalache
Dragos BurileanuDragos Burileanu
Dragos Burileanu
CBCorneliu Burileanu
Corneliu Burileanu
Speech stress detection remains an important research area, with applicability to fields and tasks such as remote monitoring, virtual assistance software, forensics operations, and even health and safety. This paper proposes a deep learning system, based on multiple Deep Neural Networks (DNNs) joined within an ensemble one-vs-one (OvO) classification strategy, using an extensive set of algorithmically extracted acoustic, prosodic, spectral, and cepstral features. The system was tested on the Speech Under Simulated and Actual Stress (SUSAS) database, for 5 class subsets and groups. Improvements have been obtained over previously reported results, with an unweighted accuracy (UA) between 62.4% and 76.1%, depending on the number of classes and their grouping.

26 July 2021
Verified
Conference Paper
Closed access
Serban MihalacheSerban Mihalache
Serban Mihalache
IIIoan-Alexandru Ivanov
Ioan-Alexandru Ivanov
Dragos BurileanuDragos Burileanu
Dragos Burileanu
In this paper, we propose a deep neural network (DNN) system for the automatic detection of speech in audio signals, otherwise known as voice activity detection (VAD). Several DNN types were investigated, including multilayer perceptrons (MLPs), recurrent neural networks (RNNs), and convolutional neural networks (CNNs), with the best performance being obtained for the latter. Additional postprocessing techniques, i.e., hysteresis thresholds, minimum duration filtering, and bilateral extension, were employed in order to boost performance. The systems were trained and tested using several data subsets of the CENSREC-1-C database, with different simulated ambient noise conditions, and additional testing was done on a different CENSREC-1-C data subset containing actual ambient noise, as well as on a subset of the TIMIT database. An accuracy of up to 99.13% was obtained for the CENSREC-1-C datasets, and 97.60% for the TIMIT dataset.

1 October 2019
Verified
Conference Paper
Closed access
Serban MihalacheSerban Mihalache
Serban Mihalache
Dragos BurileanuDragos Burileanu
Dragos Burileanu
GPGheorghe Pop
Gheorghe Pop
CBCorneliu Burileanu
Corneliu Burileanu
Speech emotion recognition is an ongoing research area, with applicability to fields such as human-machine interfaces, virtual assistance software, and forensics and law enforcement operations. Targeting the latter, a deep neural network (DNN) system is proposed for human affective state classification. The system uses modulation-based spectral features, together with the reconstruction error (RE) of the feature vector input. The RE is obtained using an autoencoder (AE), structurally identical to the DNN classifier. With this approach, improvements have been obtained over existing results on the Berlin Database of Emotional Speech (Emo-DB), reaching an unweighted accuracy (UA) of 87.3% for a reduced (5-class) set relevant for the targeted applications.

1 June 2018
Verified
Conference Paper
Closed access
GPGheorghe Pop
Gheorghe Pop
Serban MihalacheSerban Mihalache
Serban Mihalache
DBDragos Burileanu
Dragos Burileanu
The performance of speaker recognition systems based on Gaussian mixture models is often impaired both by the low quality and by the short duration of test speech samples. In literature, a large number of best material selection criteria were described, suitable for the scoring stage in forensic automatic speaker recognition systems. An application of quality-based speaker features is described in the present paper which outperforms forensic speaker recognition systems that assume uniform quality of speech during model training and scoring.

1 October 2019
Verified
Conference Paper
Closed access
Serban MihalacheSerban Mihalache
Serban Mihalache
GPGheorghe Pop
Gheorghe Pop
Dragos BurileanuDragos Burileanu
Dragos Burileanu
This work introduces the Romanian Deva Criminal Investigation Audio Recordings (RODeCAR) database, a dataset of truthful and untruthful speech, constructed by analyzing, processing, and cross-examining archived original criminal investigation recordings. The most important advantage to be leveraged when using this database is the casework nature of the content, i.e. all the speakers were suspects or witnesses in real criminal investigations, and all interactions are spontaneous and are part of actual law enforcement activity. Around 5 hours of recorded material have been extracted from archives and processed. With a total number of 19 speakers (4 female, 15 male), 39% of the content is objectively annotated as untruthful, in a global sense. Using a standard framing scheme (25 ms frames with 15 ms overlap), the 1,225,464 available segments would create a large and reliable dataset.