Automatic Detection of Parkinson’s Disease from Continuous Speech Recorded in Non- Controlled Noise Conditions J.C. Vásquez-Correa 1 , T. Arias-Vergara 1 , J.R Orozco-Arroyave 1,2 , J.F Vargas-Bonilla 1 , J.D Arias-Londoño 1 , Elmar Nöth 2 1 Faculty of Engineering, University of Antioquia UdeA 2 Pattern Recognition Lab, Friedrich Alexander Universität, Erlangen-Nürnberg Interspeech 2015
Outline 1. Introduction 2. Methodology 3. Database 4. Device 5. Results 6. Conclusion interspeech-2015 jcamilo.vasquez@udea.edu.co 2
1. Introduction Voice impairments appear in about 90% of people with Parkinson’s diasease. Reduced Loudness Only from 3% to 4% of patients recieve speech therapy Missarticulation Monotonic speech interspeech-2015 jcamilo.vasquez@udea.edu.co 3
1. Introduction Intereset: Develop computer-aided tools to perform screenings from speech. The main aims are Spare patients moving from home to the hospital Raise early alerts to patient and doctors interspeech-2015 jcamilo.vasquez@udea.edu.co 4
1. Introduction Currently The devices are mainly focused on the analysis of sustained phonations. The devices could be considered invasives. Methodologies are not adapted for evaluating speech recorded in non- controlled noise conditions. interspeech-2015 jcamilo.vasquez@udea.edu.co 5
1. Introduction interspeech-2015 jcamilo.vasquez@udea.edu.co 6
1. Introduction interspeech-2015 jcamilo.vasquez@udea.edu.co 7
1. Introduction What is missing? Portable devices and tools for the analysis of speech of people wirth Parkinson’s disease considering continuous speech signals recorded in non-controlled noise conditions. interspeech-2015 jcamilo.vasquez@udea.edu.co 8
1. Introduction Aims Portable device for recording and analysis of speech of patients Evaluation of methodology in non-controlled noise conditions. interspeech-2015 jcamilo.vasquez@udea.edu.co 9
Outline 1. Introduction 2. Methodology 3. Databases 4. Device 5. Results 6. Conclusion interspeech-2015 jcamilo.vasquez@udea.edu.co 10
2. Methodology Δ F0, Jitter, Shimmer, SVM log-Energy, Voiced MFCC Pre- V/U Speech processing segmentation signal MFCC SVM BBE Unvoiced interspeech-2015 jcamilo.vasquez@udea.edu.co 11
2. Methodology Δ F0, Jitter, Shimmer, SVM log-Energy, Voiced MFCC Pre- V/U Speech processing segmentation signal MFCC SVM BBE Unvoiced interspeech-2015 jcamilo.vasquez@udea.edu.co 12
2. Methodology Pre-processing 4 Noisy Signal x 10 Noisy Signal Frequency (Hz) 1 2 -0,5 1 Speech enhancement -1 0 0.5 1 1.5 2 0 0.5 1 1.5 2 Mean cepstral subtraction 4 x 10 Enhanced Signal Enhanced Signal Frequency (Hz) 2 0,1 0 1 -1 0 0 0.5 1 1.5 2 0.5 1 1.5 2 Time [s] Time [s] interspeech-2015 jcamilo.vasquez@udea.edu.co 13
2. Methodology Δ F0, Jitter, Shimmer, SVM log-Energy, Voiced MFCC Pre- V/U Speech processing segmentation signal MFCC SVM BBE Unvoiced interspeech-2015 jcamilo.vasquez@udea.edu.co 14
2. Methodology Segmentation Two types of sound: Voiced Unvoiced Both kind of segments are processed independently interspeech-2015 jcamilo.vasquez@udea.edu.co 15
2. Methodology Δ F0, Jitter, Shimmer, SVM log-Energy, Voiced MFCC Pre- V/U Speech processing segmentation signal MFCC SVM BBE Unvoiced interspeech-2015 jcamilo.vasquez@udea.edu.co 16
2. Methodology Characterization Features Voiced Segments Features Unvoiced Segments 12 MFCC Δ F0 25 Energy coefficients Jitter acccording to Bark Shimmer scale log-Energy 12 MFCC interspeech-2015 jcamilo.vasquez@udea.edu.co 17
2. Methodology Characterization 0.6 0.4 0.2 0 -0.2 -0.4 -0.6 -0.8 0 2 4 6 8 10 12 x 10 4 mean, 0.4 0.3 20 ms length standard deviation, 0.2 0.1 0 skewness, -0.1 Time-shift of 10 ms -0.2 -0.3 -0.4 kurtosis 0 500 1000 1500 interspeech-2015 jcamilo.vasquez@udea.edu.co 18
2. Methodology Δ F0, Jitter, Shimmer, SVM log-Energy, Voiced MFCC Pre- V/U Speech processing segmentation signal MFCC SVM BBE Unvoiced interspeech-2015 jcamilo.vasquez@udea.edu.co 19
2. Methodology Classification Gaussian kernel SVM. Parameters of the SVM are optimized in a range: 10 -1 < C < 10 4 10 -2 < γ < 10 2 Leave One Speaker Out (LOSO) Cross-validation interspeech-2015 jcamilo.vasquez@udea.edu.co 20
Outline 1. Introduction 2. Methodology 3. Database 4. Device 5. Results 6. Conclusion interspeech-2015 jcamilo.vasquez@udea.edu.co 21
3. Database Patients with Healthy controls Parkinson’s disease Num recordings 14 14 Age Mean 61.64 ± 6.43 Mean 63.29 ± 10.43 Gender 7 male, 7 female 7 male, 7 female Sampling frequency 44100 Hz Quantization bits 16 interspeech-2015 jcamilo.vasquez@udea.edu.co 22
3. Database Six different sentences One read text with 36 words interspeech-2015 jcamilo.vasquez@udea.edu.co 23
Outline 1. Introduction 2. Methodology 3. Database 4. Device 5. Results 6. Conclusion interspeech-2015 jcamilo.vasquez@udea.edu.co 24
3. Database Recording Device interspeech-2015 jcamilo.vasquez@udea.edu.co 25
Outline 1. Introduction 2. Methodology 3. Database 4. Device 5. Results 6. Conclusion interspeech-2015 jcamilo.vasquez@udea.edu.co 26
4. Results Voiced segments Sentence Accuracy (%) Noisy Enhanced 1 71 ± 26 82 ± 25 2 75 ± 26 64 ± 36 3 71 ± 26 79 ± 25 4 86 ± 31 79 ± 25 5 79 ± 25 82 ± 25 6 86 ± 23 75 ± 26 Read text 79 ± 25 71 ± 26 interspeech-2015 jcamilo.vasquez@udea.edu.co 27
4. Results Unvoiced segments Sentence Accuracy (%) Noisy Enhanced 1 92 ± 19 93 ± 17 2 94 ± 15 91 ± 20 3 86 ± 23 97 ± 12 4 93 ± 18 94 ± 16 5 78 ± 25 90 ± 20 6 86 ± 23 97 ± 12 Read text 99 ± 3 99 ± 1 interspeech-2015 jcamilo.vasquez@udea.edu.co 28
4. Results Unvoiced segments Sentence Accuracy (%) Noisy Enhanced 1 92 ± 19 93 ± 17 2 94 ± 15 91 ± 20 3 86 ± 23 97 ± 12 4 93 ± 18 94 ± 16 5 78 ± 25 90 ± 20 6 86 ± 23 97 ± 12 Read text 99 ± 3 99 ± 1 interspeech-2015 jcamilo.vasquez@udea.edu.co 29
4. Results Voiced frames Unvoiced frames Sentence 6 Sentence 6 1 1 True Positive Rate 0.8 0.8 True Positive Rate 0.6 0.6 0.4 0.4 0.2 Enhanced 0.2 Noisy 0 0 0 0.5 1 0 0.5 1 False Positive Rate False Positive Rate interspeech-2015 jcamilo.vasquez@udea.edu.co 30
Outline 1. Introduction 2. Methodology 3. Databases 4. Device 5. Results 6. Conclusion interspeech-2015 jcamilo.vasquez@udea.edu.co 31
5. Conclusion 1. A computational tool is presented for the recording and analysis of speech with Parkinson’s disease 2. The methodology evaluated considers speech recorded in non- controlled noise conditions. 3. It is useful the speech enhancement technique? interspeech-2015 jcamilo.vasquez@udea.edu.co
5. Conclusion 4. The incorporation to the methodology of prosody features derived from timing, duration, and speech rate is planned for the near future. 5. The use of the computer tool to follow the speech therapy of the patients, and to asses their neurological state is also expected in the future. interspeech-2015 jcamilo.vasquez@udea.edu.co
Thanks! interspeech-2015 jcamilo.vasquez@udea.edu.co
Recommend
More recommend
