Wim De Vilder – filter : verwijderen van ademhalingsruis uit spraaksignalen Probleemoplossen en ontwerpen, deel 3
Problem Statement Newscasters present the news with a very quick tempo Between two sentences they require a large breath Can be a distraction for the viewers Tempo can be so fast that the viewers cannot understand
Problem Statement Wim De Vilder : an example
Problem Statement Examples : Different pitch = Time-Scaling Original Signal Fast Version Slow Version Pitch Corrected
Problem Statement The Wim De Vilder-filter and Time-Scaling ? Automatically (in real-time) know the difference between speech and breath Allow speech to pass Slow down signal without distorting pitch Digital signal processing = key to the problem Characteristics (features) extracted from the acoustics Difference between speech and breath (classification) Pitch extraction from audio signal
Problem Statement The Wim De Vilder-filter and Time-Scaling ? Automatically (in real-time) know the difference between speech and breath Allow speech to pass Slow down signal without distorting pitch Digital signal processing = key to the problem Characteristics (features) extracted from the acoustics Difference between speech and breath (classification) Pitch extraction from audio signal
Problem Statement The Wim De Vilder-filter and Time-Scaling ? Automatically (in real-time) know the difference between speech and breath Allow speech to pass Slow down signal without distorting pitch Digital signal processing = key to the problem Characteristics (features) extracted from the acoustics Difference between speech and breath (classification) Pitch extraction from audio signal
Problem Statement The Wim De Vilder-filter and Time-Scaling ? Automatically (in real-time) know the difference between speech and breath Allow speech to pass Slow down signal without distorting pitch Digital signal processing = key to the problem Characteristics (features) extracted from the acoustics Difference between speech and breath (classification) Pitch extraction from audio signal
Planning Team 1 (Wim De Vilder Filter)Team 2 (Time Stretching) 30/9Problem Statement : First Group Meetings Voice activity detection : features 07/10Voice activity detection : classificatieSample rate change / framing Feature : Zero-crossing rate/periodiciteitTime Stretching : Overlap Add Synthesis (OLA) 14/10Feature : spectrale energieTime Stretching : OLA Feature : spectrale energieTime Stretching : Synchronous Overlap Add Synthesis (SOLA) 21/10Schrijven tussentijds verslag SOLA : Time Domain Auto Correlation Feature : LPCPitch Synchronous Overlap Add Synthesis (PSOLA) 25/10Deadline mid-term report 28/10Feature : cepstrale energiePitch Detection : Zero Rate Crossing 4/11Feauture : tijdsinformatiePitch Detection : Modified Zero Rate Crossing Features : combinatiePitch Detection : Auto-Correlation Techniques 12/11Bayesiaanse classificatie + Gaussian Mixture ModelsPSOLA : Implementation Bayesiaanse classificatie + Gaussian Mixture ModelsPSOLA : Implementation 18/11Real-time implementatie in Simulink 25/11Real-time implementatie in Simulink 27/11Deadline infobrochure 02/12Real-time implementatie in Simulink, preparation for demo 9/12preparation for report, presentation 16/12Presentation
Praktisch 2 sessies per week (seeTijdstabel P&O3) Monday Thursday 2 hours interaction per week for questions/problems!