Current Status and Future Challenges

1
Current Status and Future Challenges
- a Speech Technology View
Assoc. Prof. Børge Lindberg Speech and
Multimedia Communication Division (Center for
PersonKommunikation (CPK) is now integrated into
the Department of Communication
Technology) Aalborg University, Denmark E-mail
lindberg_at_kom.auc.dk
2
Current status speech recognition

• A Corpus Based Technology
• Statistical models derived from acoustic and
  linguistic databases !
• Assisted by a significant Media-development
  (storage capacity) and Hardware development
  (computational capacity)
• Appearance of Data Consortia- Linguistic Data
  Consortium (LDC), www.ldc.upenn.edu- European
  Language Resources Association (ELRA)
  www.icp.grenet.fr/ELRA/home.html
• Recently, powerful commercial speech recognition
  engines have become available

3
Current status speech recognition

• Applications in Danish
• Mainly in the telephony domain (traffic
  information, calling rates) Providers Nuance,
  Philips (now Scansoft)/PDC
• Penetration of call centre automation lower than
  in e.g. the US.
• Recently, in the office environment, also
  specialised domains such as Pathology and
  Radiology have become available more
  plannedProvider Philips/MaxManus
• No computer command control recogniser
  available for Danish and no general domain
  speech recognition available.Database will be
  soon be available from the SPEECON project
• The latter two are required for physically
  handicapped and for people having reading-,
  writing or spelling disabilities
• Other players are IBM/NST (ViaVoice), Empathy
  Systems and Scansoft (Dragon)
• Lack of robustness

4
Current status speech synthesis (TTS)

• Like ASR, a corpus based technology
• With less computational demands on the core
  engine
• Applications are out running on Danish
  SMS-to-speech, Phone-browsing, adgangforalle.dk
  (access to all), automatic generation of audio
  books
• Dedicated recorded database for target voice, -
  or larger database for trainable synthesis
• Three products available (Infovox, RealSpeak and
  DanTTS)
• Inflexible voice
• Limited prosodic control and limited NLP capacity

5
Major results achieved in Denmark

• As a result of public funding
• Text-to-Speech Synthesis is there, to a large
  extent because of public funding
• Language resources have been developed for speech
  recognition (SpeechDat family) these are
  available for research, but for a fairly high
  price
• Automatic Speech Recognition (ASR) is coming up
  (main drive is though from industry)
• Denmark is lagging behind with respect to ASR

6
What went wrong ?

• Sequential development of TTS and ASR TTS first
  ....This strategy postponed the development of
  ASR in Denmark, assisted by commercial players
  claiming they were able to do so.
• No follow-up public investment in ASR
• Databases are there but hardly accessible
• No Open Source like modules available for core
  technologies

7
Challenges and what is needed ?

• Development of no business case applications,
  mainly for disabled - public involvement, i.e.
  investment is needed
• Speech Recognisers and Text-to-speech
  synthesisers are lacking robustness and a number
  of qualities. Public involvement is needed to
  ensure continued research in new methods applied
  to the Danish language or at least to ensure
  Open-Source-like situation for core
  technologies
• Public involvement in the distribution,
  collection, validation, standardisation,
  improvement and production of language resources
  ideally freely accessible
• Without support for the smaller languages, EU
  speech technology will develop along two lines
  languages that allow full use, and languages that
  don't.
• In order to benefit from the free flow of
  researchers within EU, for that reason alone, we
  need to strive for creating and maintaining
  attractive research environments which are
  facilitating high level research
• Research into multi-modality (speech is one out
  of more modalities), multi-linguality and in
  particular non-native language.