Stanley J. Wenndt

Co-channel speaker segment separation

A novel approach to co-channel speaker separation is presented here. The technique uses the statistical properties of combinations of high Target-to-Interferer Ratio (TIR) speech segments, which were extracted from a 0 dB overall TIR co-channel utterance. The problem is broken down into making three simpler decisions. First, closed-set speaker identification technology is used on combinations of high TIR speech segments to determine which speakers are generating the co-channel speech. Next, the proportion of segments belonging to each speaker is estimated using a bimodal model. Lastly, a maximum likelihood decision is made as to which two combinations of segments best represent the two speakers. Using this approach at least one of the speakers could readily be identified when the speaker contributed a segment that was 160 ms or more in length. Once the speakers were determined, greater than 90% reliable speaker separation was obtained.

Audio steganography by amplitude or phase modification

This paper presents the results of embedding short covert message utterances on a host, or cover, utterance by modifying the phase or amplitude of perceptually masked or significant regions of the host. In the first method, the absolute phase at selected, perceptually masked frequency indices was changed to fixed, covert data-dependent values. Embedded bits were retrieved at the receiver from the phase at the selected frequency indices. Tests on embedding a GSM-coded covert utterance on clean and noisy host utterances showed no noticeable difference in the stego compared to the hosts in speech quality or spectrogram. A bit error rate of 2 out of 2800 was observed for a clean host utterance while no error occurred for a noisy host. In the second method, the absolute phase of 10 or fewer perceptually significant points in the host was set in accordance with covert data. This resulted in a stego with successful data retrieval and a slightly noticeable degradation in speech quality. Modifying the amplitude of perceptually significant points caused perceptible differences in the stego even with small changes of amplitude made at five points per frame. Finally, the stego obtained by altering the amplitude at perceptually masked points showed barely noticeable differences and excellent data recovery.

Bispectrum features for robust speaker identification

Along with the spoken message, speech contains information about the identity of the speaker. Thus, the goal of speaker identification is to develop features which unique to each speaker. This paper explores a new feature for speech and shows how it can be used for robust speaker identification. The results are compared to the cepstrum feature due to its widespread use and success in speaker identification applications. The cepstrum, however, has shown a lack of robustness in varying conditions, especially in a cross-condition environment where the classifier has been trained with clean data but then tested on corrupted data. Part of the bispectrum is used as a new feature and we demonstrate its usefulness in varying noise settings.

Blind channel estimation for audio signals

This research presents a new approach for blind channel estimation for audio signals. For most speech processing techniques such as speech recognition or speaker identification, the performance can drop significantly when the statistics of the training data such as the channel shape, noise, and distortion differ from the statistics of the testing data. Aside from the standard technique of cepstral mean normalization, few techniques are available for reducing channel mismatch conditions. Experimental results will be presented for both channel estimation and for channel normalization via inverse filtering where the inverse filter is derived from the channel estimate.

Machine recognition vs human recognition of voices

While automated speaker recognition by machines can be quite good as seen in NIST Speaker Recognition Evaluations, performance can still suffer when the environmental conditions, emotions, or recording quality changes. This research examines how robust humans are compared to machine recognition for changing environments. Several data conditions including short sentences, frequency selective noise, and time-reversed speech are used to test the robustness of both humans and machine algorithms. Statistical significance tests were completed and, for most conditions, human were more robust.

Data embedding in audio signals

This paper presents results of two methods of embedding digital audio data into another audio signal for secure communication. The data-embedded, or stego, signal is created for transmission by modifying the power spectral density or the phase spectrum of the cover audio at the perceptually masked frequencies in each frame in accordance with the covert audio data. Embedded data in each frame is recovered from the quantized frames of the received stego signal without synchronization or reference to the original cover signal. Using utterances from Texas Instruments Massachusetts Institute of Technology (TIMIT) databases, it was found that error-free data recovery resulted in voiced and unvoiced frames, while high bit-errors occurred in frames containing voiced/unvoiced boundaries. Modifying the phase, in accordance with data, led to higher successful retrieval than modifying the spectral density of the cover audio. In both cases, no difference was detected in perceived speech quality between the cover signal and the received stego signal.

Effects of cochannel speech on speaker identification

Past studies have shown that speaker identification (SID) algorithms that utilized LPC cepstral feature and a vector quantization classifier can be sensitive to changes in environmental conditions. Many experiments have examined the effects of noise on the LPC cepstral feature. This work studies the effects of co-channel speech on a SID system. It has been found that co-channel interference will degrade the performance of a speaker identification system, but not significantly when compared to the effects of wideband noise on an SID system. Our results show that when the interfering speaker is modeled as one of the speakers within he training set, it has less of an effect on the performance of an SID system than when the interfering speaker is outside the set of modeled speakers.

Narrow-band interference cancellation for enhanced speaker identification

While the cepstrum feature has been widely used for speaker identification (SID), studies have shown that it can be sensitive to changes in environmental conditions. Many experiments have examined the effects of additive white Gaussian noise on the cepstral feature, but few, if any, have been conducted using additive narrow-band interference. Since such interference appears in an unpredictable fashion due to adverse signal environments or equipment anomalies in communication systems, it is important to understand its impact along with the affect of interference removal algorithms on SID performance. This paper examines two interference removal algorithms for enhancing SID performance. One is a simple notch filter suitable for tone removal. The other is a newly introduced method suitable for mitigating more general forms of interference, including interfering signals that can be modeled as being angle-modulated.

Method and apparatus for detecting illicit activity by classifying whispered speech and normally phonated speech according to the relative energy content of formants and fricatives

Method and apparatus for the classification of speech signals. Speech is classified into two broad classes of speech production—whispered speech and normally phonated speech. Speech classified in this manner will yield increased performance of automated speech processing systems because the erroneous results that occur when typical automated speech processing systems encounter non-typical speech such as whispered speech, will be avoided.

Content recognition for telephone monitoring

This research began due to federal inmates abusing their telephone privileges by committing serious offenses such as murder, drug dealing, and fraud. On average, about 1000 calls are made per day at each federal prison with a peak of over 4000. Current monitoring capabilities are very man- intensive and only allow for about 2-3% monitoring of inmate telephone conversations. One of the main deficiencies identified by prison officials is the need to flag phone conversations pertaining to criminal activity. This research looks at two unique voice-processing methods to detect phone conversion pertaining to criminal activity. These two methods are digit string detection and whisper detection.