Speaker dependency

 

The system may be designed to tolerate a large variety of speaker variabilities. In this case the system is speaker independent   and has to deal with a large population of users, mainly from the general public .

Other systems may be tuned to the voice of a particular speaker (a single speaker) and thus the system is speaker-dependent.   We may also have a system that is adapted to the voice of a particular set of speakers (multi-speaker system ).

Some systems may also be adapted to be utilised only by trained speakers instead of the general public (untrained speakers). Others may be used by consistent or inconsistent speakers, skilled users/unskilled users, a specific ``population'', or the general public . Some may be used, for a given language, by native speakers or by speakers with foreign or heavy (regional) accents .

In order to achieve speaker independency a training phase  is usually needed and uses a specified type of speech corpus. The characteristics of the required speech data have to be described in terms of

• type of data
  • speech acoustic wave forms,
  • acoustic data with phonetic labelling ,
  • acoustic data with the corresponding orthographic forms,
  • acoustic data with the corresponding phonetic transcription  ,
  • acoustic data with the corresponding recognition-units transcription, etc.)
• size of data (how many hours/minutes of speech),
• number of speakers and how they are selected
  • arbitrary, or
  • with respect to some characteristics such as
    • sex ,
    • age  or age ranges,
    • physical state , psychological,
    • experience,
    • attitude,
    • accent ,
    • etc.
• acquisition channels 
  • single microphone ,
  • set of microphones,
  • similar telephone handset, or
  • as many handsets as possible,
• environment  conditions (noisy , quiet, all conditions, etc.) and
• many other constraints derived from the operating condition.