This dissertation describes a number of algorithms developed to increase the robustness of automatic speech recognition systems with respect to changes in the environment. These algorithms attempt to improve the recognition accuracy of speech recognition systems when they are trained and tested in different acoustical environments, and when a desk-top microphone (rather than a close-talking microphone) is used for speech input. Without such processing, mismatches between training and testing conditions produce an unacceptable degradation in recognition accuracy.
Two kinds of environmental variability are introduced by the use of desk-top microphones and different training and testing conditions: additive noise and spectral tilt introduced by linear filtering. An important attribute of the novel compensation algorithms described in this thesis is that they provide joint rather than independent compensation for these two types of degradation.
Acoustical compensation is applied in our algorithms as an additive correction in the cepstral domain. This allows a higher degree of integration within SPHINX, the Carnegie Mellon speech recognition system, that uses the cepstrum as its feature vector. Therefore, these algorithms can be implemented very efficiently. Processing in many of these algorithms is based on instantaneous signal-to-noise ratio (SNR), as the appropriate compensation represents a form of noise suppression at low SNRs and spectral equalization at high SNRs.
The compensation vectors for additive noise and spectral transformations are estimated by minimizing the differences between speech feature vectors obtained from a "standard" training corpus of speech and feature vectors that represent the current acoustical environment. In our work this is accomplished by minimizing the distortion of vector-quantized cepstra that are produced by the feature extraction module in SPHINX.
In this dissertation we describe several algorithms including the SNR-Dependent Cepstral Normalization, (SDCN) and the Codeword-Dependent Cepstral Normalization (CDCN). With CDCN, the accuracy of SPHINX when trained on speech recorded with a close-talking microphone and tested on speech recorded with a desk-top microphone is essentially the same obtained when the system is trained and tested on speech from the desk-top microphone.
An algorithm for frequency normalization has also been proposed in which the parameter of the bilinear transformation that is used by the signal-processing stage to produce frequency warping is adjusted for each new speaker and acoustical environment. The optimum value of this parameter is again chosen to minimize the vector-quantization distortion between the standard environment and the current one. In preliminary studies, use of this frequency normalization produced a moderate additional decrease in the observed error rate.

/pdf/acoustical-and-environmental-robustness-in-automatic-speech-3yf73aaxj5.pdf

Acoustical and environmental robustness in automatic speech recognition

With the advent of `multimedia', digital signal processing (DSP) of sound has emerged from the shadow of bandwidth limited speech processing to become a research field of its own. To date, most research in DSP applied to sound has been concentrated on speech, which is bandwidth limited to about 4 kilohertz. Speech processing is also limited by the low fidelity typically expected in the telephone network. Today, the main applications of audio DSP are high quality audio coding and the digital generation and manipulation of music signals. They share common research topics including perceptual measurement techniques and analysis/synthesis methods. Additional important topics are hearing aids using signal processing technology and hardware architectures for digital signal processing of audio. In all these areas the last decade has seen a significant amount of application-oriented research. The frequency range of wideband audio has an upper limit of 20 kilohertz and the resulting difference in frequency range and Signal to Noise Ratio (SNR) due to sample size must be taken into account when designing DSP algorithms. There are whole classes of algorithms that the speech community is not interested in pursuing or using. These algorithms and techniques are revealed in this book. This book is suitable for advanced level courses and serves as a valuable reference for researchers in the field. Interested and informed engineers will also find the book useful in their work.

Applications of digital signal processing to audio and acoustics

This thesis addresses the problem of how a listener groups together acoustic components which have arisen from the same environmental event, a phenomenon known
as auditory scene analysis. A computational model of auditory scene analysis is
presented, which is able to separate speech from a variety of interfering noises.
The model consists of four processing stages. Firstly, the auditory periphery is
simulated by a bank of bandpass filters and a model of inner hair cell function. In
the second stage, physiologically-inspired models of higher auditory organization -
aiditory maps - are used to provide a rich representational basis for scene analysis.
Periodicities in the acoustic input are coded by an ant ocorrelation map and a crosscorrelation map. Information about spectral continuity is extracted by a frequency
transition map. The times at which acoustic components start and stop are identified
by an onset map and an offset map.
In the third 8tage of processing, information from the periodicity and frequency
transition maps is used to characterize the auditory scene as a collection of symbolic auditory objects. Finally, a search strategy identifies objects that have similar
properties and groups them together. Specifically, objects are likely to form a group
if they have a similar periodicity, onset time or offset time.
The model has been evaluated in two ways, using the task of segregating voiced
speech from a number of interfering sounds such as random noise, "cocktail party"
noise and other speech. Firstly, a waveform can be resynthesized for each group
in the auditory scene, so that segregation performance can be assessed by informal
listening tests. The resynthesized speech is highly intelligible and fairly natural.
Secondly, the linear nature of the resynthesis process allows the signal-to-noise ratio
(SNR) to be compared before and after segregation. An improvement in SNR is
obtained after segregation for each type of interfering noise. Additionally, the performance of the model is significantly better than that of a conventional frame-based
autocorrelation segregation strategy.

/pdf/computational-auditory-scene-analysis-a-representational-10sb27ppe0.pdf

Computational auditory scene analysis : a representational approach

The technique fits a sinusoidal model to additive vocal speech segments so that the least-mean-squared error between the model and the summed waveforms is obtained. Enhancement is achieved by synthesizing a waveform from the sine waves attributed to the desired speaker. Least-squares estimation is applied to obtain sine-wave amplitudes and phases of both talkers, based on either a priori sine-wave frequencies or a priori fundamental frequency contours. When the frequencies of the two waveforms are closely spaced, the performance is significantly improved by exploring the time evolution of the sinusoidal parameters across multiple analysis frames. The least-squared-error approach is also extended, under restricted conditions, to estimate fundamental frequency contours of both speakers from the summed waveforms. The results obtained, although limited in their scope, provide evidence that the sinusoidal analysis/synthesis model with effective parameter estimation techniques offers a promising approach to the problem of cochannel talker-interference suppression over a range of conditions. >

An approach to co-channel talker interference suppression using a sinusoidal model for speech

The improvement of identification accuracy of concurrent vowels with differences in fundamental frequency (ΔF0) is usually attributed to mechanisms that exploit harmonic structure. To decide whether identification is aided primarily by selecting the target vowel on the basis of its harmonic structure (‘‘harmonic enhancement’’) or removing the interfering vowel on the basis of its harmonic structure (‘‘harmonic cancellation’’), pairs of synthetic vowels, each of which was either harmonic or inharmonic, were presented to listeners for identification. Responses for each vowel were scored according to the vowel’s harmonicity and that of the vowel that accompanied it. For a given target, identification was better by about 3% for a harmonic ground unless the target was also harmonic with the same F0. This supports the cancellation hypothesis. Identification was worse for harmonic than for inharmonic targets by 3%–8%. This does not support the enhancement hypothesis. When both vowels were harmonic, identificatio...

/pdf/identification-of-concurrent-harmonic-and-inharmonic-vowels-46k1chm121.pdf

Identification of concurrent harmonic and inharmonic vowels: A test of the theory of harmonic cancellation and enhancement

The problem addressed by this study is the suppression of an undesired talker when two talkers are communicating simultaneously on the same monophonic channel (co-channel speech). Two different applications are considered, improved intelligibility for human listeners, and improved performance for automatic speech and speaker recognition (ASR) systems. For the human intelligibility problem, the desired talker is the weaker of the two signals with voice-to-voice power ratios (Power desired / Power interference), or VVRs, as low as -18dB. For ASR applications, the desired talker is the stronger of the two signals, with VVRs as low as 5dB. Signal analysis algorithms have been developed which attempt to separate the co-channel spectrum into components due to the two different (stronger and weaker) talkers.

J. Naylor

Papers

Techniques for suppression of an interfering talker in co-channel speech