[Column] Artificial Intelligence IV

-
Voice and AI (transform music to data)
  1. speech and music
  2. cock-tail party effect
  3. 85-1100 Hz for Human
  4. Steps from Human voice to MP3 
    1. VOICE > SENSOR > VOLTAGE
    2. SAMPLING (computer cannot manage continuous signal) : DISCRETE (time)
    3. QUANTIZATION : DISCRETE (amplitude)
    4. ENCODING
  5. TIME SERIES
  6. WAVEFORM
  7. SAMPLING RATE (for MP3 is 44100 Hz , the reason is human is not sensitive to high frequency voice )
Characteristics of Voice
  1. frequency spectrum (x - frequency ; y - amplitude) with log coordinate
  2. Amplitude (Loudness)
  3. Frequency (Pitch)
  4. Timbre (related with harmonics)
Classic Voice Features -MFCC
  1. MFCC - Mel-Frequency Cepstral Coefficients
    1. Formant
    2. Vowel
  2. Steps of using MFCC
    1. 26 dimensions vector 
      1. Mel-Frequency, mel(f) = 1125 *ln(1+f/700)
      2. good classification at low frequency , bad at high frequency (similar as human)
    2. Calculate Its CEPSTRAL
      1. the purpose of cepstral is converting 26 dimensions into 13 dimensions.
      2. window width
      3. window pitch
    3. finally get 13 dimensions MFCC features
  3. Deep learning applied (feature extract and feature classify)
    1. convolutional layer (convert into much more specific feature)
    2. pooling layer  (reduce dimension)
    3. full connected layer (combined convolution by inner product calculation)
    4. softmax layer (outcome probability)
Application - SPEECH RECOGNIZATION
  1. voice assistant (without type writing , meeting record)
  2. acoustic model
  3. language model
  4. music automatically searching ! (fuzzy searching) - window scan

Comments

Post a Comment

Popular posts from this blog

[SW] LMTD Calculations

-
LMTD Calculations LMTD Calculations [Maintenance: BingHung ; Ver: 20180611_R0] Item Unit Nomenclature Value Hot Side Temperature In oC Thin Hot Side Temperature Out oC Thout Cold Side Temperature In oC Tcin Cold Side Temperature Out oC Tcout LMTD Calculation - Co-current flow dT1 oC dT2 oC LMTD - Co-Current Flow oC LMTD Calculation - Counter-current flow dT1 oC dT2 oC LMTD - Counter-Current Flow oC
Read more

[Column] Artificial Intelligence I & II

-
AI Development History 1950 : Alan Turing , Thesis - Computing Machinery and Intelligence , Turning Test 1951 : Marvin Minsky, Princeton Master Student, SNARC (1969 Turning Award) 1955 : Allen Newell/ Herbert Simon/ Cliff Shaw, Logic Theorist, Reasoning    1956 : John McCarthy/ Claude Shannon/ Nathan Rochester, AI Artificial Intelligence First Period (1956-1974) 1963 : APPA, Project MAC, MIT CSAIL 1964-1966, Joseph Weizenbaum, ELIZA, MIT 1967-1972, Human Robot. Japan Second Period (1980-1987) 1960s, Edward Feigenbaum, expert system 1970s, Stanford, MYCIN 1980s, CMU,DEC,XCON,expert system Expert system Artificial neural network 1982, John Hopfield, Hopfield net, Associative memory 1986, David Rumelhart/ Geoffrey Hinton/ Ronald Williams, Thesis-Learning representations by back-propagating errors, prove backpropagation by experiment 1982,Japan, 5g calculator, over ten years investment, the project finally fail. Third Period (2011-Now) Math Based and S...
Read more

[Column] Artificial Intelligence III

-
Recap :What are AI good at ? Classification ! Two steps of classification : feature extract and feature classify What is image mean to computer ? pixel , matrix , resolution  Black and white , 0 - 255 Colorful (0-255,0-255,0-255), RGB , tensor  Channel number (# dimension) Tensor  0 order : scalar 1 order : vector 2 order : matrix Before deep learning (human designed feature) graphical feature is defined by computer vision, edge texture after combined with machine learning  object recognization object detection Convolution vector, matrix, tensor can be applied. steps of convolution : repeat (slide, extract, inner product) convolution kernel  histogram of oriented gradients (HOG) - to get the edge of figure  step 1 : using convolution calculation get the edge of the figure  step 2 : divide figure into multiple zones  step 3 : according to edge feature , amplitude, orientation conduct statistic.  step 4 :...
Read more