Goals of the Project

The goals for the PUMS project in Joensuu are the following.

1. Implement state-of-the-art speaker recognition technique
   • Implement standard signal processing methods that are needed in creation of speaker profiles and in speaker identification, e.g.
     • Cepstrum computation,
     • Delta-coefficients,
     • Cepstral Mean Subtraction algorithm,
     • Optimally speaker discriminating filter bank (instead of e.g. mel-spectrum filter bank),
   • Management of speaker profiles created from speech signals
   • Classification methods, for example, implement Gaussian Mixture Model classifier to complement the currently used VQ classifier
2. Port the speaker recognition software in the project partners' application platforms
   • PC software for speaker recognition and profile management, and
   • a speaker recognition application running in a mobile phone
3. Systematically test the created software applications and investigate the recognition quality
   • in different acoustic conditions,
   • with different equipment.
4. Long term target: develop novel methods for the speaker recognition task
   • Fusion of several speaker cues, including low-level acoustic cues and prosodic cues
   • Segmentation methods for automatic tagging of discriminative speech segments (discriminative feature extraction)
   • Development of discriminative training and matching algorithms
   • Confidence measures for speaker recognition
   • Development of efficient algorithms for real-time use

Goals in Detail

The project goals are described in more detail in the internal project plan. The parts of PUMS project that are developed in University of Joensuu deal with partial projects

B2.

Improvement of the Off-line Speaker Recogniser

B4.

Speaker Recognition and Speaker Profiles

The B2 goals consist of

• Improve stability of the present software
• Improve usability
• Develop better likelihood estimates
• Automatic computation of weights for the committee classification
• Integrating on-line and off-line recognisers in the same software
• Automation of phonetic features

The B4 goals consist of

• Automatic speaker recognition in a mobile phone
• Speaker recognition tool library and Windows software utilizing it