System Identification Related Problems at SMN

Transcription

1 Ericsson research SeRvices, MulTimedia and Networks System Identification Related Problems at SMN Erlendur Karlsson SysId Related ER/SMN Ericsson External Page 1

2 Outline Research Ericsson Research System Identification related applications at SMN Important issues when dealing with real-world problems SysId Related ER/SMN Ericsson External Page 2

3 Research Ericsson Research Ericsson Research Blogg 5G Cloud Context Aware Communication Data and Knowledge Internet of Things LTE Media Coding SDN Security Sevice Systems Smart Cities SysId Related ER/SMN Ericsson External Page 3

4 Contextual Communiction Excavator MWC 2015 Excavator Excavators from Volvo CE Powerful Linux PC Python application with custom signaling built on top of OpenWebRTC Control Rig Simulator from Oryx and Volvo CE Mac OS X computer OS X Cocoa application with custom signaling built on top of OpenWebRTC Signaling Server SysId Related ER/SMN Ericsson External Page 4

5 Remote Excavation Technologies Spatial scene capture, both video and audio Spatial scene rendering, both video and audio Low latency real time communication Low latency remote control SysId Related ER/SMN Ericsson External Page 5

6 Media Processing Architecture Camera Microphone Array Excavator Audio Video Data OpenWebRTC Audio / Video Data Channel Audio Video Data Network Mobile, fixed, Network Audio Video Mobile, fixed, Data OpenWebRTC Audio / Video Data Channel Audio Video Data Oculus Rift DK2 Headphones Control Rig SysId Related ER/SMN Ericsson External Page 6

7 System Identification Related Applications at MMT Audio and Speech Coding Audio Mining (ASR) Audio Media Processing Acoustic Echo Cancellation Noise Suppression Voice Activity Detection Spatial Audio Capture Spatial Audio Rendering Video Coding (2D and 3D) Objective Quality Estimation of Encoded Audio and Video Congestion Control in IP Networks SysId Related ER/SMN Ericsson External Page 7

8 Audio and Speech Coding Clean speech signals can be modeled very efficiently with Code-Excited Linear Prediction (CELP) encoders (Based on ARX model of the speech signal) Music signals are better encoded with transform encoding methods (Subband filter banks, MDCT) Signal classification and hybrid encoding used to obtain efficient encoding of audio signals of varying content EVS (Enhanced Voice System) just standardized in 3GPP standardization Special EVS session at ICASSP 2015 in Australia SysId Related ER/SMN Ericsson External Page 8

9 CELP Speech Model SysId Related ER/SMN Ericsson External Page 9

10 Bitstream Bitstream EVS Speech/Audio Codec prototype HL structure Mode TD TD- BWE Improved AMR-WB technology Parametric high band Technology Linear Pred. + ACELP FCB variable sf. Linear prediction, energy/gain FD G.719-like Transform (LD-MDCT), block switching input VAD Mode Dec. TD (+TD-BWE) FD CNG wb WB SWB FB bandwidths TD AMRWB-like TD-BWE FD G.719 like FD-coding parametric 4 ~ Audio BW [khz] SysId Related ER/SMN Ericsson External Page 10

11 Acoustic Echo Cancellation Long echo impulse reponses: msec At 48 khz sampling : 14,400 24,000 samples SysId Related ER/SMN Ericsson External Page 11

12 Spatial Audio Capture Microphone arrays Filter design in the spatial and frequency domains Beamforming techniques Adaptive tracking of the most active speakers in a room SysId Related ER/SMN Ericsson External Page 12

13 Spatial Audio Rendering Spatial hearing 3D binaural rendering through Head Related Filtering (HRF) Very useful in 3D gaming and evolved communication solutions Spatial audio rendering onto any loudspeaker configuration SysId Related ER/SMN Ericsson External Page 13

14 Spatial Hearing SysId Related ER/SMN Ericsson External Page 14

15 Acoustic Wave Reception The listeners median plane Sound wave Left Head Related Filter (HRF) Contralateral ear Listener Right Head Related Filter (HRF) Ipsilateral ear Length L ITD = L/c where c=speed of sound SysId Related ER/SMN Ericsson External Page 15

16 ASR System Main Components Training Data Acoustic Models Applying Lexical Models Constraints Language Models Speech Signal Representation Feature Vector Search Recognized Words Speech recognition is the problem of deciding on How to represent the signal How to model the constraints How to search for the most optimal answer SysId Related ER/SMN Ericsson External Page 16

17 ASR System Solution Components Acoustic- Phonetic Modeling Pattern Recognition Finite-State Transducers Language Models Adaptation Acoustic Models Lexical Models Language Models Speech Signal Representation Search Recognized Words Speech Signal Representation Search Algorithms Vector Quantization & Clustering Hidden Markov Modeling Graphical Models Segmental Models GMMs Neural Networks SysId Related ER/SMN Ericsson External Page 17

18 Important issues when dealing with real-world problems Understand the strengths and weaknesses of the different identification methods Preprocessing the data before the optimization can be crucial Choose the minimization criterion with care and adapt it to the problem at hand Different type of regularization components in the criterion can make the difference between success and failure Some times a criterion having components in both the time and frequency domains will work, when single domain criterions fail. SysId Related ER/SMN Ericsson External Page 18

19 Important issues when dealing with real-world problems Some applications require classification based modelling, where the current model used depends on signal classification of some signals Many systems have to deal with spurious events This will require the detection of such events and special model updates when they are detected Monitoring of system model Hypothesis testing and estimation SysId Related ER/SMN Ericsson External Page 19

20 Erlendur Karlsson,