In this paper we present a method for echo detection and echo path delay estimation using a pattern recognition approach. We consider the problem of echo detection as attempting to match a speech pattern in the near-end signal to the far-end signal at a given delay. Employing features and techniques that have been successfully used in speech recognition, we define a spectral similarity function based on cepstral correlation. We show, through experimental results, that the proposed similarity function can reliably detect acoustic echoes and correctly estimate the echo path delay. Further, it is shown that the similarity function can be used in the detection of double-talk conditions. The method presented here is applicable to both electrical (hybrid) network echoes as well as to acoustic echoes.

In new generation networks like 3G wireless and VoIP, a great deal of emphasis is put on transcoder-free operation (TrFO), where speech remains coded throughout the core network. Any network-based speech processing function must, therefore, operate on the coded parameters directly if the value of TrFO is to be realized. Many of these functions, like echo control, noise reduction, and gain control can be viewed as dynamic amplitude scaling of the speech signal. Given that an intermediate step of decoding/re-encoding is not an option, we present, in this paper, a method for dynamic scaling of speech in the coded-domain directly. We derive expressions for modifying the relevant coded parameters such that the resulting decoded speech would correspond to the desired scaled signal. Experimentally, we use the AMR 12.2 kbps coder, and show that the proposed method results in a signal whose level, as well as speech quality, closely matches the desired scaled signal

Compared to the landline network environment, the wireless environment presents new factors affecting ASR performance (or accuracy). Our goal here is to determine these factors, and their relative importance, and then to devise methods to mitigate them. We approach this goal, first, by conducting a set of experiments where we use a state of the art ASR system trained on landline speech data, and then compare its performance in landline network conditions to its performance across a variety of wireless network conditions. Based on the results of these experiments, we determine critical factors affecting ASR accuracy. We then use multi-condition acoustic models to mitigate these factors and show that the resulting ASR system is able to not only achieve high accuracy across the various wireless network conditions, but also maintain its high accuracy across landline network conditions. This leads to a recognition system that is channel independent which is a very desirable property for telecom based applications.

In this paper, we present a method that manipulates the decoding network to reduce both computational complexity and response latency while maintaining high ASR accuracy. The method employs a TSVQ (tree structured vector quantization) classifier that reliably discriminates between silence and non-silence frames. Reductions in computational complexity and response latency are achieved through three techniques: 1) silence skipping, 2) silence-based pruning of the dynamic programming network, and 3) early decision. Experimental results on a connected digit task and a large vocabulary company name task show that the proposed method can reduce ASR response latency by more than 82%. Furthermore, the computational complexity, measured in CPU seconds, was reduced by 13.6% on the connected digit task and 6.7% on the company name task while maintaining the recognition accuracy of the baseline system.

A new approach for speaker verification is presented. Mixture decomposition discrimination (MDD) is based on the idea that, when modeling speech using speaker independent continuous density hidden Markov models (HMM), different speakers speaking the same word would cause different HMM mixture components to dominate. When the mixture information is considered, one can construct a "mixture profile" of a speaker speaking a given word or phrase. This mixture profile is incorporated into a discriminative training procedure to discriminate between a true speaker and all other speakers (or imposters). The effectiveness of MDD is seen when it is incorporated into a hybrid verification system that also includes speaker dependent HMM modeling with cohort normalization. Experimental results show that the hybrid system reduces the average equal error rate (EER) by 46% when compared with the EER of the speaker-dependent HMM verifier. It is also shown that the computational and model storage requirements needed to incorporate MDD into the hybrid system are relatively small.

Vector quantization (VQ) has been explored in the past as a means of reducing likelihood computation in speech recognizers which use hidden Markov models (HMMs) containing Gaussian output densities. Although this approach has proved successful, there is an extent beyond which further reduction in likelihood computation substantially degrades the recognition accuracy. Since the components of the VQ frontend are typically designed after model training is complete, this degradation can be attributed to the fact that VQ and HMM parameters are not jointly estimated. In order to restore the accuracy of a recognizer using VQ to aggressively reduce computation, joint estimation is necessary. We propose a technique which couples VQ frontend design with minimum classification error training. We demonstrate on a large vocabulary subword task that in certain cases, our joint training algorithm can reduce the string error rate by 79% compared to that of VQ mixture selection alone.

We formulate a training framework and present a method for task independent utterance verification. Verification-specific HMMs are defined and discriminatively trained using minimum verification error training. Task independence is accomplished by performing the verification on the subword level and training the verification models using a general phonetically balanced database that is independent of the application tasks. Experimental results show that the proposed method significantly outperforms two other commonly used task independent utterance verification techniques. It is shown that the equal error rate of false alarms and false keyword rejection is reduced by more than 22% compared to the other two methods on a large vocabulary recognition task.

In this paper, we present a new approach for speaker verification, termed mixture decomposition discrimination (MDD). MDD is based on the idea that, when modeling speech using speaker independent continuous density hidden Markov models (HMM), different speakers speaking the same word would activate different HMM mixture components. One can therefore construct a "mixture profile" of a speaker speaking a given word or phrase. This mixture profile is incorporated into a discriminative training procedure to discriminate between a true speaker and imposters. The effectiveness of MDD is seen when it is incorporated into a hybrid verification system that also includes speaker dependent HMMs with cohort normalization. Experimental results show that the hybrid system reduces the average equal error rate (EER) by 52% when compared with a stand alone cohort-normalized speaker dependent HMM verifier. It is also shown that the computational and model storage requirements needed to incorporate MDD into the hybrid system are relatively small.

An utterance verification method based on minimum verification error training is presented. In a two-stage process, the recognition hypothesis produced by an HMM-based speech recognizer is verified using a set of verification-specific models that are independent of the models used in the recognition process. The verification models are trained using a discriminative training procedure that seeks to minimize the verification error by simultaneously maximizing the rejection of non-keywords and misrecognized keywords while minimizing the rejection of correctly recognized keywords. This method is evaluated on a connected digit recognition task with a null grammar. The baseline string error rate for this task was 4.85%. At 5% rejection of valid strings, the string error rate decreased to 2.70% using the proposed verification method. The corresponding performance on non-keyword speech was a rejection rate of over 99.0%.

An integral part of any deployable speech recognition system is the capability to detect if the input speech does not contain any of the words in the recognizer vocabulary set. This capability, which is called utterance verification (or keyword recognition and nonkeyword rejection), is therefore becoming increasingly important as speech recognition systems continue to migrate from the laboratory to actual applications. We present a framework and a method for vocabulary independent utterance verification in subword-based speech recognition. The verification process is cast as a statistical hypothesis test, where vocabulary independence is accomplished through a two-stage verification process: subword-level verification followed by string-level verification. A verification function is defined and discriminatively trained to perform subword-level verification. String-level verification is accomplished by defining and evaluating an overall string-level log likelihood ratio that is a function of the subword-level verification scores. Experimental results show that this vocabulary-independent discriminative utterance verification method significantly outperforms a baseline method commonly used in wordspotting tasks.

Connected digit recognition is an area that has attracted significant attention because of its importance in automating a wide range of applications. In an actual application, the performance of the connected digit recognizer must be measured along two dimensions: its recognition accuracy over connected digit strings, and its rejection rate over utterances with no digits. This paper presents a rejection method for connected digit recognition. This rejection method is a post-processor to an HMM-based recognizer and consists of two stages: a digit/non-digit classification stage, and a string verification stage. The digit/non-digit classifier is discriminatively trained to determine if a given digit segment in the recognized string actually contains a digit. The string verification stage combines the results of the digit/non-digit classifier to make the final rejection decision. Experimental results on two independent databases, show that this rejection method is successful in not only rejecting speech with no connected digits, but also in rejecting putative errors, which would have resulted in misrecognition.>

An iterative method for simultaneous pole-zero parameterization of a given filter impulse response is presented. An error criterion is constructed and iteratively minimized. The error criterion emphasizes both spectral peaks and spectral valleys in the desired signal spectrum and results in a set of linear equations that can be readily solved. A comparison with well-known methods (e.g. Shanks' and Kalman's methods) shows that this simultaneous iterative method is substantially better. Experimental results also show that a significant reduction in error is achieved after only three or four iterations. Applications of this method include system identification and pole-zero modeling of speech.>

The design and implementation of a parallel-processing-based pitch detector is presented. Pitch information is extracted by performing pitch detection on four different waveforms derived from the speech signal. Pitch information from the four pitch-detection processes is then combined to determine a final pitch estimate. The performance of this pitch detector is evaluated on a large database and compared to other well-known pitch detection algorithms. It has been implemented in real time on a TMS32020 fixed-point digital signal processor as part of a 2.4 kb/s vocoder. A performance comparison of the real-time fixed-point implementation and a computer simulation are also given. The results show that the pitch detector performance is maintained in the real-time implementation mainly because the majority of the algorithm computations are integer arithmetic and logic-type operations.>