SIGNAL PROCESSING TOOLS FOR SPEECH RECOGNITION

1
SIGNAL PROCESSING TOOLS FOR SPEECH RECOGNITION
Presented by Richard Duncan
Hualin Gao, Richard Duncan, Julie A. Baca, Joseph
Picone Human and Systems Engineering Center of
Advanced Vehicular System Mississippi State
University
2
INTRODUCTION
The tools described here deal with the block
known as the Acoustic Front-end, which
encapsulates the signal processing portions of a
recognition system.
3

• THE PROBLEM WITH AVAILABLE TOOLS

Why reinvented the wheel? The problem with
existing tools

• Run-time efficiency
• File I/O
• Adding new algorithms

The goal of this paper is to solve these
problems.
4

• FEATURES OF ISIP FOUNDATION CLASSES

• Unicode support for multilingual applications
• Memory management and tracking
• System and I/O libraries that abstract users from
  details of the operating system
• Math classes that provide basic linear algebra
  and efficient matrix manipulations
• Data structures that include generic
  implementations of essential tools for speech
  recognition code.

5

• DESIGN REQUIREMENTS FOR THESE TOOLS

• A library of standard algorithms to provide basic
  digital signal processing (DSP) functions
• An ability to easily add new algorithm classes
  and functions without modifying existing classes
• A block diagram approach to describing algorithms
  to realize rapid prototyping without programming.
  
• A mechanism to share the code base for the
  feature extraction and recognizer.

6

• BASIC DIGITAL PROCESSING FUNCTIONS

This example shows how to realize the basic
digital signal processing functions. It computes
the energy of input vector in dB using the SUM
algorithm // declare an Energy object and an
output vector // Energy egy VectorFloat
output // set the input vector // VectorFloat
input(L"0, 1, 2") // choose algorithm //
enrgy.setAlgorithm(EnergySUM) // choose
implementation // egy.setImplementation(Energy
DB) // compute the energy of input data //
egy.compute(output, input)
7

• ADDING NEW ALGORITHMS

• The implementation using an abstract base class,
  AlgorithmBase, and virtual functions or methods
  that comprise the interface contract, is the
  single most important feature, since it makes the
  library extensible.
• All algorithm classes are derived from this base
  class.

8
HIERARCHY OF ALGORITHM CLASSES
New classes can be easily added by following the
interface contract
9

• SIGNAL PROCESSING CONFIGURATION TOOLS

The menu of the tools
10

• SIGNAL PROCESSING CONFIGURATION TOOLS

Create recipes and configure each block by right
click it
11
BLOCK DIAGRAM TO DESIGN FRON-TEND

• Users specify a pool of recipes for use in the
  Front-End.
• Each recipe is an information holder for
  Algorithm objects.
• The Front-End creates a dependency graph. In the
  top left recipe pool, recipe 6 can not process
  before recipe 5 has been processed.
• Processing tool runs a breadth first search to
  produce output from given inputs.

12

• THE SIGNAL PROCESSING TOOLS

• Uses a popular block diagram approach for
  designing signal processing algorithms.
• Allows rapid prototyping of ideas.
• Allows users to easily integrate new modules into
  the IFCs and have them instantly available in all
  tools, such as the recognizer.

13
THE SIGNAL PROCESSING TOOLS
The signal processing control tool is a driven
program and is called isip_transform.exe in our
environment. It uses the signal processing
library and algorithm libraries and output recipe
files from signal configuration tool to fulfill
the whole procedure of signal processing.
14
THE SIGNAL PROCESSING TOOLS

• Parsing the file containing the recipe created by
  the user with the configuration tool
• Synchronizing different paths along the block
  flow diagram contained in this file
• Preparing input/output data buffers for each
  algorithm, particularly for those requiring
  multiple frames of data, such as windows or
  calculus
• Scheduling the sequences of required signal
  processing operations
• Processing data through the flow defined by the
  recipe
• Managing conversational data.

15
THE SIGNAL PROCESSING TOOLS
The processing procedure
16
MFCC TWO PASSES RECIPES
?First pass to get the cepstrum mean subtraction
and maximum energy normalization.
Second pass to ? get the feature values by using
the results from the first pass.
17

• TEST THE TOOLS

• The correctness The implementation of each
  algorithm is verified manually or by using other
  tools such as MATLAB.
• Usability Assessed and enhanced the usability of
  our tools through extensive user testing
  conducted over the course of many workshops.
• Speech recognition experiments The correctness
  of the tools was also proved by several speech
  recognition experiments.

18

• CONCLUSIONS

• A library of standard algorithms for basic DSP
  functions.
• The ability to add new algorithms to this library
  easily.
• A GUI-based configuration tool for creating block
  diagrams to describe algorithms, allowing rapid
  prototyping without programming.
• Tested and verified these tools for both
  correctness and usability.