Text to Speech System for the Hindi Language

1
Text to Speech System for the Hindi Language

• Project Team
• Anupam
• Brijesh
• Guide Dr. Amitabh Mukherjee

2
Formato de la Presentación

• Introduction
• Motivation
• Architecture
• Work already done in the field
• Problems encountered
• Our approach

3
Introducción

• Text-To-Speech (TTS) is a process through which
  input text is
• step 1 analyzed
• step 2 processed
• step 3 understood
• Intonation and prosody
• , then the text is rendered as
  digital audio and then spoken.

4
Motivación

• In the movie 2001 A Space Odyssey A computer
  HAL talks like a human being. It was a brilliant
  imagination in 1960s. Now this dream has come
  true for English what for hindi movies ???? .
• Dependence of human computer interaction on
  written texts and images, makes the use of
  computers impossible for visually and physically
  impaired and illiterate masses
• TTS has not been extensively dealt with in the
  Hindi Domain

5
La Arquitectura
Source CDAC home page
6
Continue

• Grapheme to Phoneme mapping
• Speech synthesis

7
Advantages and problems with Hindi

• Hindi resembles Sanskrit. Deterministic rules
  for a language like Sanskrit - which is called
  loyal - in the sense that  people speak what they
  write. .
• Schwa deletion The effect of not pronouncing
  a in certain positions as in the terminal a
  for Himalaya.
• Anusvara ambiguity This is caused because the
  grapheme for anusvara can represent any nasal
  sound. eg aMbara ( sky ) , aMda ( egg )

8
Previous Works

• Grapheme to Phoneme mapping
• 1. Media lab in IIT kgp
• Rule based Grapheme to Phoneme mapping
  by Monojit choudhary
• Approach
• three category C Half
• V Full not the matras
  
• CV Full
• Observations from our day-today speaking , few
  rules are made to delete schwa and for Anuswara
  and chandra-bindu .

9
Example

• The schwa of first syllable is never deleted
• Eg kalama ( pen ) , Badara (cloud )
• Our Approach
• DFA is constructed
• to delete schwa
• To get proper phoneme for chandra-bindu and
  answara string .

10
Speech Synthesis

• Quality
• Intelligibility measured with phonemes,
  syllables, words, phrases
• Naturalness Difficult to get
• 1. Formant In this kind of TTS Systems
  Voice is generated by the simulation of the
  behaviour of human vocal cord . Robotic voice
  
• 2. Concatenation syllables are
  concatenated at run time and they produce
  phonetic representation of text. large storage
  
• 3. Sinusoidal speech signal can be
  represented as a sum of sine waves with time
  varying amplitude and frequencies.

11
Work already done

• CDAC has used concatenation method to develop
  natural voice
• Anuvacchaka
• Uchharaka
• Bhavna
• Bilss technology hybridization of Formant
  and Concatenated. We called it RecSimCat .

12
ANN approach

• neural networks are trained from actual speech
  samples, they have the potential to generate more
  natural sounding speech than other synthesis
  technologies.
• ANN structure
• First ANN
• Generation of segment durations from
  phonetic descriptions
• Second ANN
• Generation of acoustic
  information from phonetic and timing information
  speech parameters

13
Speech Synthesis with Neural NetworksOrhan
Karaali, Gerald Corrigan, and Ira Gerson
14

• Contacting IITkgp to know more about this .

15
Des Questions?
16
Merci beaucoup!