Can MARF detect Chinese tonal differences

  • Hi

    I have a voice recognition project and I am looking for programmer who can help me do the job.

    I posted the job on and one coder mentioned MARF. That's how I get to know MARF.

    My project is a personal Chinese dictation program. It is speaker-dependent meaning that each user has to record 1300 distinctive syllables as the 'standard' against which future dictation should refer to as the standard.

    This project is,therefore, much easier than a speaker-independent one.

    But most coders that were interested in my project were unable to differentiate the 4 tones of Chinese sound syllables. It's because in most languages (like Arabic), the tones plays no part in the language. So they are unfamiliar to process the Chinese sounds that depend heavily on tones to differentiate between syllables.

    My question is: Can MARF detect the subtle differences of the 4 tones of the same sound? If it can, then I'd like to know how I can find coders that are familiar with MARF to do my project.


    • Hi,

      My belief that can be possible with some of the algorithm
      combinations in MARF (not all would work, only frequency-based ones
      for feature extraction).

      This is an interesting project; I said 'belief' because no-one
      has actually tried Chinese tones yet for this task. But setting
      up such an experiment should be easy enough with the current MARF
      as-is provided the needed samples data is available and modification
      to the speakers.txt in SpeakersIdentApp file accordingly to the
      data samples. Since the tones are based on the frequency changes,
      the change of frequency "features" should be easy enough to detect.

      I could set this up for you fast if you could share your sample
      recordings, or just give the training and testing sample file
      names, and we can produce the required speakers.txt. Each entry
      in this file would be the tone you are trying to identify, with the
      accompanying samples to train and to test on. You will require
      MARF itself as well as the SpeakerIdentApp for this task. I would
      recommend getting the one from the CVS instead of the FRS, as those
      are somewhat behind the current state of things.

      Let me know if you need more info.