This thesis presents methods and results for optimizing subword detectors in continuous speech. Speech detectors are useful within areas like detection-based ASR, pronunciation training, phonetic analysis, word spotting, etc. Firstly, we propose a structure suitable for subword detection. This structure is based on the standard HMM framework, but in each detector the MFCC feature extractor and the models are trained for the specific detection problem. Our experiments in the TIMIT database validate the effectiveness of this structure for detection of phones and articulatory features. Secondly, two discriminative training techniques are proposed for detector training. The first one is a modification of Minimum Classification Error training. The second one, Minimum Detection Error training, is the adaptation of Minimum Phone Error to the detection problem. Both methods are used to train HMMs and filterbanks in the detectors, isolated or jointly. MDE has the advantage that any detection performance criterion can be optimized directly. F-score and class accuracy optimization experiments show that MDE training is superior to the MCE-based method. The optimized filterbanks reflect some acoustical properties of the detection classes. Moreover, some changes are consistent over classes with similar acoustical properties. In addition, MDE-training of filterbanks results in filters significatively different than in the standard filterbank. In fact, some filters extract information from different critical bands. Finally, we propose a detection-based automatic speech recognition system. Detectors are built with the proposed HMM-based detection structure and trained discriminatively. The linguistic merger is based on an MLP/Viterbi decoder.

Zugl.: München, Techn. Univ., Diss., 2009

Thesis--Brunswick. / In Periodical Room.

Zugl.: Erlangen, Nürnberg, University, Diss., 2009.

Thesis--RWTH Aachen, 2000.

Zugl.: Karlsruhe, Univ., Diss., 2008

Techn. University, Diss., 2003--Berlin.

Wireless systems with multiple antennas at both the transmitter and receiver (MIMO systems) have been the focus of research in the recent past due to their ability to provide higher data rates and better reliability than their single antenna counterparts. Designing a communication system for MIMO frequency selective channels provides many signal processing challenges. Popular methods like MIMOOFDM and space-time precoding linearly process blocks of data at both the transmitter and the receiver. Independence between the blocks is ensured by introducing sufficient redundancy between successive blocks. This approach has many pitfalls, including the limit on achievable data rate due to redundancy requirements and the need for additional coding/processing. In this thesis, we provide a filterbank precoding framework (FBP) for communication over MIMO frequency selective channels. By viewing the channel as a polynomial matrix, we derive the minimum redundancy required for achieving FIR equalization of the precoded channel. It is shown that, for most practical channels, a nominal redundancy is enough. The results are general, and hold for channels of any dimension and order. We derive the zero-forcing and MMSE equalizers for the precoded channel. The role of equalizer delay in system performance is analyzed. We extend the minimum redundancy result to the case of space-time filterbank precoding (STFP). Introducing the time dimension allows the channel to be represented by a block pseudocirculant matrix. By using the Smith form of block pseudocirculant matrices, we show that very high data rates can be achieved with STFP. When channel information is available at the transmitter, we derive an iterative algorithm for obtaining the MMSE optimal precoder-equalizer pair. We then provide a comparison of FBP with the block processing methods. It is shown that FBP provides better BER performance than the block processing methods at a lower computational cost. The reasons for the better performance of FBP are discussed.

<p>The report descripbs how a diplexer for a hybrid analog/digital filterbank has been constructed and tested. A diplexer divides the frequency band into two different bands that do not who doesn't overlapp each other. The sampling rate for the two ADC:s is 80 Msps, and therefore it is advantage to have zero at 80 MHz. The reason for this is that a proposed class of hybrid filterbanks with very good quality requires a zero at or close the sampling frequency to work well. </p><p>The diplexer was made in three versions. The first didn't work since the choosen inductance self resonance frequency was to low and by the same range as the filters bandwidth. The second version had to much losses, which resulted in attenuation at 80 MHz, which was to small. The third version was made in two differents layout. </p><p>To the diplexer it was also made a test tool in Labview, through that one gets the magnitude ande phase functions. </p><p>The results show that the magnitude function is good for version 3.0 and for version 3.1, and that the losses are low. It depends also on that the choosen components have a high self resonance frequency. There can't been shown any differences between these two, thus is is difficult to judge if one is better then an other.</p>

The report descripbs how a diplexer for a hybrid analog/digital filterbank has been constructed and tested. A diplexer divides the frequency band into two different bands that do not who doesn't overlapp each other. The sampling rate for the two ADC:s is 80 Msps, and therefore it is advantage to have zero at 80 MHz. The reason for this is that a proposed class of hybrid filterbanks with very good quality requires a zero at or close the sampling frequency to work well. The diplexer was made in three versions. The first didn't work since the choosen inductance self resonance frequency was to low and by the same range as the filters bandwidth. The second version had to much losses, which resulted in attenuation at 80 MHz, which was to small. The third version was made in two differents layout. To the diplexer it was also made a test tool in Labview, through that one gets the magnitude ande phase functions. The results show that the magnitude function is good for version 3.0 and for version 3.1, and that the losses are low. It depends also on that the choosen components have a high self resonance frequency. There can't been shown any differences between these two, thus is is difficult to judge if one is better then an other.