You are on page 1of 11

Hindi Speech to Text

Conversion

Vivek Kumar Pandey


JIIT , Noida (INDIA)
Recent Developments in
Speech Recognition in
Indian
1.CDAC-NOIDA Languages
A-STAR/U-STAR Project
2.CIIL-LDCIL
3.KIIT- Mobile Text & Speech Database Collection in
Hindi and Indian Spoken English
(Contracted by Nokia Research Center, China)
POS Tagging

 Different words in a sentence can be classified into


different categories called Parts of Speech.
e.g. NN,VB,ADJ,ADV,PREP,PRO,DET etc.
 Different Types of taggers
1.Rule based
2.Stochastic
3.Hybrid
Statistical Approaches

 Histogram Analysis
 Hidden Markov Model (HMM)
 Maximum Entropy Model (ME)
 Conditional Random Field (CRF)
 Memory Based Learning (MBL)
POS Tagger : Basic Requirements

 Tag Set


Corpus
Indian Languages : Tag Sets
Available


Very few tag sets available publicly


IL Tagset -IIT Hyderabad : very coarse structure in
linguistic analysis, resulting into a very flat structure
e.g. Tag “PREP” used for POSTP also.
Corpus Generation

 Unicode Supported Factors deciding accuracy :


 Collected from some  Number of sentences in
small stories , novels or the corpus
newspaper articles  Tokens/word
Framework : Our Approach

 Statistical approache : require a huge


training set
 Pattern Matching : Rely on native speakers (Hindi in
our case)
 Database design
 Aids to improve accuracy : Huge database and
machine learning
Future Work

 Can be implemented for other Indian languages by


making changes in the database.

 Size of the Corpus and number of tokens can be


increased to improve accuracy
References

 Speech and Language Processing by Daniel Jurafsky & H. Martin



Natural Language Processing and Information Retrieval by Tanveer
Siddiqui & U.S. Tiwari
 Hindi Word Sense Disambiguation by Manish Sinha , Mahesh Kumar
Reddy .R , Pushpak Bhattacharyya , Prabhakar Pandey & Laxmi
Kashyap
 Cryptanalysis of Keystream Reuse in Stream Ciphered Digitized
Speech using HMM based ASR Techniques by L. A. Khan and M.S.
Baig

You might also like