Segmentation and Classification of Vowel Phonemes of Assamese Speech Using a Hybrid Neural Framework

Algorithm 1

SOM weight vector extraction algorithm.

(1) Input: Spoken word 𝑆 of size 𝑚×𝑛, sampling

    frequency 𝑓_𝑠, duration 𝑇 second

(2) Preprocess the signal using preprocessing

  algorithms described in Section 7.1

(3) Initialize 𝑃=20, order of linear prediction

(4) Find the coefficients of a 𝑃th-order linear predictor FIR filter,

  𝑄=−𝑎(2)𝑄(𝑛−1)−𝑎(3)𝑄(𝑛−2)−⋯−𝑎(𝑝+1)𝑄(𝑛−𝑝)

   that predicts the current value of the real-valued

   preprocessed time series 𝑄, based on past samples.

(5) Store 𝑎=[1,𝑎(2),𝑎(3),…,𝑎(𝑝+1)]

(6) Take an topology map whose neurons are arranged in

  an 1×(𝑃+1)-dimensional hexagonal pattern.

(7) Initialize weight 𝑊_𝑖(𝑘) to a small random number.

(8) Initialize learning parameter, 𝜂 and neighbors, ℵ_𝐼((𝑘)

(9) for 𝑘=1 to (𝑝+1) do

    pick a(k)

  find Wining neuron as,

 ‖𝑎(𝑘)−𝑊_𝑖(𝑘)‖=min_{(1≤𝑖≤(𝑝+1))}{‖𝑎(𝑘)−𝑊_𝑖(𝑘)‖}

 Update synaptic vectors of the wining cluster,

   𝑊_𝑖(𝑘)=𝑊_𝑖(𝑘)+𝜂_𝑘(𝑎(𝑘)−𝑊_𝑖(𝑘))

        Update 𝜂_𝑘, ℵ_𝐼((𝑘)

(10) Store updated weight 𝑊_𝑖(𝑘) as SW_𝑗, where 𝑗=1∶6