SpeechLab - Speech Verification System Overview
Uploaded by
AmrSpeechLab - Speech Verification System Overview
Uploaded by
AmrTeam Members Amr Medhat
Speech Lab
An approach to teaching the Holy Quran recitation rules
Mostafa Fathi Sameh Serag Supervised by Dr. Nevin Darwish
THE MAIN IDEA
Speech Lab Lab Speech
Users Speech
Corrected Pronunciation
Introduction
Pronunciation Teaching: The Problem
One-to-many Distributed feedback Harder error-detection Repetition Time consuming Not available all the time
Teaching) (Computer-Assisted Pronunciation
The rise of CAPT
Speech Lab: The Solution
Objective:
CAPT system for teaching the Holy Qur'an recitation rules.
Why?
The same problems of pronunciation teaching The market need Serves Islam and Muslims
Previous Work
ISLE
Interactive Spoken Language Education (ISLE) project, funded by the European Commission English for adult Italian and German learners
LISTEN
research project at Carnegie Mellon University Automated reading tutor for children
Idea
System Learner's speech Reference speech Feedback
Recognition of learners speech Compare with a reference speech Provide appropriate feedback
How?
By detecting phone insertions, deletions and substitution.
Phone (or phoneme): The smallest sound unit of speech
e.g. phone substitution in the word
a_l
start
n
a_h
s:
end
What can we do?
Different recitation rules require different way in handling This approach can handle rules like
Our model handles ( ) for 8 letters.
Design
2
Phone-level
HMM-based
Utterance saved
1
Users Utterance Feedback
10
GUI
Utterance File
Recognizer
Feedback
Recognized word 4
Feedback Generator
Users mistakes
User Profile Analyzer
Pronunciation difference
String Comparator
Mistakes filtered
7
Recognized word compared with reference word
Building Recognizer
Building Speech Database Creating Transcriptions Signal Analysis (Coding)
Decoding
Constructing Recognition Lattice
Training HMMs
Analyzing Results
Development Tools
HTK (HMM Tool Kit)
free toolkit for research in ASR and other pattern recognition systems developed at Cambridge University. consists of a set of modules in C for building Hidden Markov Models .
Cool Edit Pro Microsoft Visual Studio .NET
Challenges
Lack of resources on CAPT Arabic phonology and Quran Phonology
Speech database
Time Sheet
Task Research Learning the tool Experiments Design & Prototype Implementation
GUI & Integration & Testing
Sept
Oct
Nov
Dec
Jan
Feb
Mar
Apr
May
Jun
Jul
Documentation
Thank you
QUESTIONS ??
For contact: [email protected]
HMM
HMM (contd)
Model Parameters {}
Transition probabilities A ={ aij } Observation probabilities B ={bj }
In speech:
Observations represent frames of 10 ms States usually correspond to parts of a phone
Parameter estimation: Baum-Welch Decoding: Viterbi P (O | )
Signal Analysis
