Speech Enhancement Group

Head of SubGroup:

Doc. Ing. Pavel Sovka, CSc.

Permanent members:

Ing. Petr Pollak, CSc.

Doc. Ing. Vratislav Davidek, CSc.

Ing. Pavel Masa

PHd students:

Ing. Jiri Sika

Ing. Milan Svoboda

Students:

Jan Kybic

The basic areas of interest are mentioned with brief description. More detailes can be found in the published articles which are also available on-line.

SPEECH ENHANCEMENT FOR COMMUNICATION PURPOSES.

The research is focused to enhancement of speech in the running car environment. That means the additive noise model can be assumed. The presented solutions can be used in many other applications if the assumption of additive noise is fulfilled. We have studied the following solutions:

one channel algorithms based on standard spectral subtraction
The spectral subtraction is the most simple algorithm of noise suppression in the frequency-domain. The different approaches were studied and compared. The optimal case of spectral subtraction from the point of view 'musical noise' minimalization were suggested.
Persons: Pavel Sovka | Petr Pollak | Milan Svoboda
Published works:

1. The Noise Suppression System for a Car. - EUROSPEECH'93
   

extended spectral subtraction
The extended spectral subtraction is modification of adaptive Wiener filtering where Wiener filter is used for estimation of noise which is then subtracted from corrupted speech signal. The transfer function of Wiener filter is evaluated from the characterisitics computed during previous frame processing.
Persons: Pavel Sovka | Petr Pollak
Published works:

1. Extended Spectral Subtraction. - EUSIPCO'96
   

twochannel and multichannel noise suppression systems
Multichannel system for speech enhancement in the running car were tested. Some experiments were realized with 4 microphone array. The following problems were solved:
- placement of microphones
- time-delay estimation and compensation for different channels
- noise suppression in the 4-channel system
Persons: Pavel Sovka | Pavel Masa

experiments with Kalman filtering
This solution is currently an important poit of interest. One student starts working on his diploma project in this field. But the results comparable to other above presented algorithms are not available in this time.
Persons: Petr Pollak | Jan Kybic

VOICE ACTIVITY DETECTORS - (VAD).

Many one channel systems estimate the noise characteristics in the speech pauses. The Voice Activity Detector (VAD) has to be involved in such a system. In our noise suppression systems we have used according to target application energy based or especially cepstral VAD.
Persons: Petr Pollak | Pavel Sovka
Published works:

1. Cepstral Speech/Pause Detectors. - IEEE Workshop on Nonlinear Signal and Image Processing, 1995.
   
2. The Study of Speech/Pause Detectors for Speech Enhancement Methods. - EUROSPEECH'95.
   

ROBUST SPEECH RECOGNITION.

the combination of speech enhancement techniques and standard recognition systems
The robust recognition system derived from the isolated word recognizer realized in HTK were tested. The above described algortihms for speech enhancement were used in the pre-processing step. Generally, the recognition rate can be improved this way. This improvement is proportional to the improvement of the SNR of noisy speech signal. But the distortion of enhanced speech yields little further decreasing of finally achieved recognition rate.
Persons: Petr Pollak | Pavel Sovka | Tomas Kreisinger
Published works:

1. Study of Speech Recognition in Noisy Environment. - ECSAP'97.
   

the robust parametrization techniques
We have studied also the direct robust parametrization techniques for speech corrupted by additive noise. The first experiments were realized on corellation subtraction technique implemented into standard LPC parametrization.
Persons: Petr Pollak | Pavel Sovka
Published works:

1. Robust LPC Parametrization Based on Noise Correlation Compensation. (PostScript)
   

THE REAL-TIME IMPLEMENTATION OF ABOVE DESCRIBED ALGORITHMS.

Some algorithms were implemented into signal processor of family TMS320C30. The algorithms of spectral subtraction with different modifications and different VADs were implemented for processing in the real-time. The emulator module on PC card with signal processor TMS320C30 were used in these experiments. More details are on Real-Time Homepage.
Persons: Vratislav Davidek | Jiri Sika

BACK TO
Czech Technical University | Faculty of Electrical Engineering | Department of Circuit Theory
Speech processing group

This page is always under construction! Please, mail me any suggestion or problem. Contact: pollak@feld.cvut.cz Last change: 20 Mar 1997