SLIDE 1
TTS and Data Selection: Improving Systems for Low-Resource Languages
Chevy Levitan, DREU 2015
- utline
- I. Project
- II. Approach
- III. Methods
- IV. Status
- V. Future
TTS and Data Selection: Improving Systems for Low-Resource Languages - - PowerPoint PPT Presentation
TTS and Data Selection: Improving Systems for Low-Resource Languages - - PowerPoint PPT Presentation
TTS and Data Selection: Improving Systems for Low-Resource Languages Chevy Levitan, DREU 2015 outline I. Project II. Approach III. Methods IV. Status V. Future I. Project synthesize natural, intelligible voices for low resource languages
SLIDE 2
SLIDE 3
I.
Project
synthesize natural, intelligible voices for low resource languages using data selection
SLIDE 4
motivation
▷ bridge the gap
SLIDE 5
motivation
▷ bridge the gap ▷ allow for cross-language communication
SLIDE 6
why data selection?
SLIDE 7
HRLs vs. LRLs
★ prepared data ★ abundance of training material high quality speech systems ★ found data ★ limited training material low quality speech systems
SLIDE 8
- A. filter out unwanted data from
training set
SLIDE 9
- A. filter out unwanted data from
training set
- B. supplement limited LRL data with
choice data from similar HRL
SLIDE 10
II.
APPROACH
preparing the experiment
SLIDE 11
▷ Boston Radio News Corpus ▷ pre-processed ▷ English
corpus
SLIDE 12
data selection process
extract features sort values create subsets synthesize data
SLIDE 13
evaluate.
SLIDE 14
evaluate.
compare/contrast voices
SLIDE 15
example
VOICE 1 VOICE 2
SLIDE 16
solution
- 1. subset data
- 2. complete dataset
SLIDE 17
III.
METHODS
testing our hypothesis
SLIDE 18
★ follow standard procedures for evaluating TTS voices
standards
SLIDE 19
★ follow standard procedures for evaluating TTS voices ★ successful voice = intelligible + natural
standards
SLIDE 20
★ follow standard procedures for evaluating TTS voices ★ successful voice = intelligible + natural ★ use crowdsourcing for unbiased results
standards
SLIDE 21
Intelligibility
➔ transcribe nonsense sentences ➔ accurate transcription = intelligible voice
mechanical turk
SLIDE 22
Intelligibility
➔ transcribe nonsense sentences ➔ accurate transcription = intelligible voice
mechanical turk
Naturalness
➔ use Likert scale to rate voices from very unnatural to very natural ➔ identify the voices are categorized as natural+
SLIDE 23
SLIDE 24
IV.
STATUS
- ur current state
SLIDE 25
✓ create subsets
intelligibility HIT
SLIDE 26
✓ create subsets ✓ synthesize voices with this data
intelligibility HIT
SLIDE 27
✓ create subsets ✓ synthesize voices with this data ✓ design and implement HIT
intelligibility HIT
SLIDE 28
✓ create subsets ✓ synthesize voices with this data ✓ design and implement HIT ✓ publish on MTurk site
intelligibility HIT
SLIDE 29
✓ create subsets ✓ synthesize voices with this data ✓ design and implement HIT ✓ publish on MTurk site ✓ workers complete HITs
intelligibility HIT
SLIDE 30
✓ created subsets ✓ synthesized voices with this data ✓ design and implement HIT ✓ publish on MTurk site ✓ workers complete HITs ✓ accept/reject work
intelligibility HIT
SLIDE 31
✓ create subsets
naturalness HIT
SLIDE 32
✓ create subsets ✓ synthesize voices with this data
naturalness HIT
SLIDE 33
✓ create subsets ✓ synthesize voices with this data ✓ design and implement HIT
naturalness HIT
SLIDE 34
✓ create subsets ✓ synthesize voices with this data ✓ design and implement HIT
- publish on MTurk site
- workers complete HITs
- accept/reject work
naturalness HIT
SLIDE 35
V.
FUTURE
further exploration of this research
SLIDE 36
evaluation
analyze mechanical turk responses
SLIDE 37
evaluation
analyze mechanical turk responses
low-resource
implement data selection for LRLs
SLIDE 38
evaluation
analyze mechanical turk responses
text
apply similar methods to automatically select text data
low-resource
implement data selection for LRLs
SLIDE 39