Fist page

1
Country Report Vietnam Luong Chi Mai Institute
of Information Technology Vietnamese Academy of
Science and Technology lcmai_at_ioit.ac.vn
2
VLSP National Project (ICT Program)

• National Project 2006 2008, 2008 2010 with
  participation of ten research groups (all active
  groups on VLSP)
• Objectives
• Basic research on methods for processing
  Vietnamese language and speech
• Build and develop several typical products for
  VLSP for public end-users.
• Build and develop indispensable resources and
  tools for the VLSP development

3
Objective of the Project Basic research

• Basic research on methods for processing
  Vietnamese language and speech.
• Applied research to adapt methods, technologies,
  advanced techniques for other languages to
  Vietnamese language and speech.

Typical products for the end-users
Resources and tools for VLSP
Computation methods for VLSP
4
Phonetic Structure
Pitch
(7)
(1)
(5)
(3)
(6)
(2)
(4)
(8)
Time

• (7) and (8) have F0 contour similar to (5) and
  (6), but rise and fall more sharply
• (8) is not accompanied by glottalization

5
Some Current Text Corpora

• Monolingual corpora VLC (Vietnam Lexicography
  Centre), UNS-VNUHCM, etc. for Vietnamese
• Bilingual corpora The EVC corpus (UNS-VNUHCM)
  consists of 400,000 pairs of E-V sentences
  (approx. 5,500,000 words) in the fields of
  Science and Technology (Computer,
  Electronics,..). This EVC has been being
  partially annotated with morphology (word
  boundary, lemmatize), POS and Sense tags
  semi-automatically.

6
Some Current Speech Corpora

• Broadcasting Speech Corpus VOV (Voice of Vietnam)
  contains ? 23,000 utterances, ? 4,000 distinct
  syllables
• 30 broadcasters and speakers reciting stories,
  news reports, colloquy
• Data digitized at 16,000 Hz sampling rate, using
  16 bits per sample
• All data were manually transcribed at syllable
  level
• At phonetic level Corpus contains all Vietnamese
  phonemes, but is not phonetically balanced (up to
  50 capacity from story reading programs)
• The number of speakers is limited and most
  speakers are Northern persons. So, corpus does
  not cover most variations of Vietnamese speech

7
Some Current Speech Corpora

• Telephone Speech Corpus
• Mobile phone 170 speakers from the North (males
  55) and (females 45), 1600 digit strings
• Cordless phone 208 speakers from the South (130
  males, 78 female, 442 utterances with 2340 words
• Labeling at syllable level, labeled manually at
  phonetic level, using forced alignment with
  manual adjustment (using HTK and CSLU toolkit)
• Develop Dialog System for Continuous Digit
  Recognition and VnTTS for reading SMS on Smart
  phone (Symbian).

8
Some Current Speech Corpora

• Vietnamese TTS Speech corpus
• one female voice from a short story
• 567 utterances of an average length of 15
  syllables, about 40.000 syllables
• 11 kHz sampling rate, and 16-bit resolution
• Corpus is labeled in syllable level, segment
  boundaries
• Prosody detection (Fujisaki model) CART for
  manipulation of duration
• Develop Vietnamese TTS system VnVoice based on
  PSOLA

9
Some Current Speech Corpora

• VNSpeech corpus
• 5 different kinds of units Phoneme, Tones,
  Digits and string of digits, application words,
  sentences and paragraphs
• text collected by using a web-robot (about 2500
  websites in Vietnam) with about 10,020,000
  sentences.
• 50 speakers
• Sentences corpus is divided into two parts, a
  common part and a private part
• The common part 33 conversations and 37
  paragraphs. They were read by all speakers.
• The private part about 2,000 short paragraphs,
  each speaker was asked to read 40 paragraphs.

10
Some Current Speech Corpora

• Distribution of mono-phones in speech corpus and
  Web corpora

11
Some Current Speech Corpora

• Distribution of six tones in speech corpus and
  Web corpora

12
Design New Corpora

• Main goal design and realize of corpora
  available
• to provide the Vietnamese researchers with a
  basic amount of speech material for general
  speech research, including speech synthesis and
  speech recognition
• for developing commercial speech recognition
  engines in given purposes (number recognizer,
  limited command recognizer, name recognizer, ...)

13
Design of General Corpus

• General Corpora
• for general purpose to do research on continuous
  independent-speaker recognition with the large
  vocabulary
• selected text from text corpus, and text selected
  by linguistics
• Number of speakers 200-300 (50 -male and 50 -
  female), ages 15-45
• Number of sentences 300 sentences, each sentence
  is spoken by one speaker for at lease 3 times.
• Size of vocabulary 3000-4000 Vietnamese
  syllables
• Requirement context balance is obtained among
  Vietnamese phonemes

14
Design of Specific Corpus

• Continuous Digit Corpus
• is used to make the recognition applications
• Number of speakers 100-200 (50 -male and 50 -
  female), ages 15-45
• Concurrence of digits should be approximately the
  same for all digits, the sentences consist of
  digits with random order and have variant
  lengths.
• The number of word will be from 10000 words, each
  sentence is recorded at least 3 times for each
  speaker.
• Name corpus
• Popular Vietnamese names including family name
  and first name.
• Number of speaker 100-200 (50 -male and 50 -
  female), ages 15-45
• Each sentence consists of one full name and
  spoken by a speaker for at least 3 times.
• The sentence should contain as many as possible
  names of people in Vietnam. The size of
  vocabulary is estimated about 2000 words.

15
Corpus Organization

• The database consists of several sub-databases
• 1. /general Material for training, testing the
  general recognizer
• 2. /Digit Material for training, testing digit
  continuous recognizer
• 3. /Name Material for training, testing proper
  name recognizer
• .... Other sub-database to be inserted
  depending on the specific purposes
• Each sub-database has directory hierarchy is as
  follows
• 1. / database root directory
• 2. /train to be used for system training
• 3. /test material to be used for system
  testing
• 4. / doc online documentation and tables
  unusual)
• The train and test directories contain
  sub-directories corresponding to each speaker,
  whose names are coded as follows XXXSRR where
• XXX speaker identifier
• S sex code F for female, M for male
• RR region code (HN for Hanoi, SG for Saigon,
  HU for Hue).
• Each sentence directory contains 3
  sentence-related files.
• wav wave file
• phn - phonetically-based transcription

16
Thank you !