http://blog.csdn.net/zouxy09/article/details/9156785/http://www.zhihu.com/question/20962240
Pipeline1. Collect speech and silence (aka Background ambient noise) Audiocollector-record audio data and save it as raw audio file.-if The volume of the audio recording is either too high or too low,Adjust the audio gain by tweaking M recordinggain in Microphonerecorder.java Adjust the audio gain by tweaking M recordinggain in Microphonerecorder.java. 2.audacity-label the Data 3.audiofeatureextractor-read Raw Audio file from the collector and a label file from Audacity to obtain a "features . Arff "file is used by Weka to generate a speech detection classifier 4.once You has generated your speech detection classifier you need to build the actual signal processing pipeline wit H your new classifier. You need to add your pipeline into the audio collector, or your Lab 2 app.
The MFCC computation pipeline (
mel‐frequency spectral coefficients)1.window:20-40ms, characterized by relative stability2.Power Spectrum: Calculates the Power Spectrum of each frame to identify the primary frequency in each frame3.Apply Mel Fliterbank: The higher the frequency, the more difficult to distinguish the similar frequency; combine the frequency spectrum into bins that's similar to how our ear perceive Analog Human EarThe first
filter is very narrow and gives a indication of how much energy exists near 0 Hertz where human hearing is Very sensitive to variations. As the frequencies get higher our filters get wider as we become less concerned about variations.Covert from frequency to Mel scale4. Logarithm of the Mel Filterbank. Loudness and Energy logarithmic relationship5. DCT of the log filter Bank:there is a lot of correlations between the log Filterbank energies, and this step tri Es to extract the most useful and independent features.
MFCC coefficients; Together, you get a-a-element acoustic vector that is the core features used in speech processing algorithms.
Other audible features:characteristics of sound, eg. prosody1.Pitch2.Intensity3.Temporal Aspects4.Voice Quality(glottal waveform)5.spectrogram:describe the energy distribution across frequency bands (speak dependent and emotion related)
Speech Processing
identify the combination of phonemesUse a Hidden Markov Model (HMM).a HMM for each of the alphabet tries to look for a sequence of phonemes
diagnose for mental illness Gaussian Mixture Models (GMM): A Clustering approacheach cluster can be described with a potential Gaussian distribution.
monitoring affect with a Mobile Phone1.Cellphone Monitoring of healthy subjects as part of a healthcare package2.Subsidized "Callingcard" number for Atrisk populations3.Monitoring of Humancomputer speech interfaces and interpersonal speech for elders in assisted or independent care 4.Monitoring Stress in everyday lives5.Monitoring Social Interactions (or lack of it)
Speech recognition-background