Speech coding | EPFL Graph Search

Speech coding is an application of data compression to digital audio signals containing speech. Speech coding uses speech-specific parameter estimation using audio signal processing techniques to model the speech signal, combined with generic data compression algorithms to represent the resulting modeled parameters in a compact bitstream. Common applications of speech coding are mobile telephony and voice over IP (VoIP). The most widely used speech coding technique in mobile telephony is linear predictive coding (LPC), while the most widely used in VoIP applications are the LPC and modified discrete cosine transform (MDCT) techniques. The techniques employed in speech coding are similar to those used in audio data compression and audio coding where appreciation of psychoacoustics is used to transmit only data that is relevant to the human auditory system. For example, in voiceband speech coding, only information in the frequency band 400 to 3500 Hz is transmitted but the reconst

Multi-parametric source-filter separation of speech and prosodic voice restoration

Olaf Schleusing

In this thesis, methods and models are developed and presented aiming at the estimation, restoration and transformation of the characteristics of human speech. During a first period of the thesis, a concept was developed that allows restoring prosodic voice features and reconstruct more natural sounding speech from pathological voices using a multi-resolution approach. Inspired from observations with respect to this approach, the necessity of a novel method for the separation of speech into voice source and articulation components emerged in order to improve the perceptive quality of the restored speech signal. This work subsequently represents the main part of this work and therefore is presented first in this thesis. The proposed method is evaluated on synthetic, physically modelled, healthy and pathological speech. A robust, separate representation of source and filter characteristics has applications in areas that go far beyond the reconstruction of alaryngeal speech. It is potentially useful for efficient speech coding, voice biometrics, emotional speech synthesis, remote and/or non-invasive voice disorder diagnosis, etc. A key aspect of the voice restoration method is the reliable separation of the speech signal into voice source and articulation for it is mostly the voice source that requires replacement or enhancement in alaryngeal speech. Observations during the evaluation of above method highlighted that this separation is insufficient with currently known methods. Therefore, the main part of this thesis is concerned with the modelling of voice and vocal tract and the estimation of the respective model parameters. Most methods for joint source filter estimation known today represent a compromise between model complexity, estimation feasibility and estimation efficiency. Typically, single-parametric models are used to represent the source for the sake of tractable optimization or multi-parametric models are estimated using inefficient grid searches over the entire parameter space. The novel method presented in this work proposes advances in the direction of efficiently estimating and fitting multi-parametric source and filter models to healthy and pathological speech signals, resulting in a more reliable estimation of voice source and especially vocal tract coefficients. In particular, the proposed method is exhibits a largely reduced bias in the estimated formant frequencies and bandwidths over a large variety of experimental conditions such as environmental noise, glottal jitter, fundamental frequency, voice types and glottal noise. The methods appears to be especially robust to environmental noise and improves the separation of deterministic voice source components from the articulation. Alaryngeal speakers often have great difficulty at producing intelligible, not to mention prosodic, speech. Despite great efforts and advances in surgical and rehabilitative techniques, currently known methods, devices and modes of speech rehabilitation leave pathological speakers with a lack in the ability to control key aspects of their voice. The proposed multiresolution approach presented at the end of this thesis provides alaryngeal speakers an intuitive manner to increase prosodic features in their speech by reconstructing a more intelligible, more natural and more prosodic voice. The proposed method is entirely non-invasive. Key prosodic cues are reconstructed and enhanced at different temporal scales by inducing additional volatility estimated from other, still intact, speech features. The restored voice source is thus controllable in an intuitive way by the alaryngeal speaker. Despite the above mentioned advantages there is also a weak point of the proposed joint source-filter estimation method to be mentioned. The proposed method exhibits a susceptibility to modelling errors of the glottal source. On the other hand, the proposed estimation framework appears to be well suited for future research on exactly this topic. A logical continuation of this work is the leverage the efficiency and reliability of the proposed method for the development of new, more accurate glottal source models.

EPFL2012

Optimal segmentation techniques for piecewise stationary signals

Paolo Prandoni

The concept of stationarity is central to signal processing; it indeed guarantees that the deterministic spectral properties of linear time-invariant systems are also applicable to realizations of stationary random processes. In almost all practical settings, however, the signal at hand is just a finite-size vector whose underlying generating process, if we are willing to admit one, is unknown; In this case, invoking stationarity is tantamount to stating that a single linear system (however complex) suffices to model the data effectively, be it for analysis, processing, or compression purposes. It is intuitively clear that if the complexity of the model can grow unchecked, its accuracy can increase arbitrarily (short of computational limitations); this defines a tradeoff in which, for a given data vector, a set of complexity/accuracy pairs are defined for each possible model. In general one is interested in parsimonious modeling; by identifying complexity with "rate" and model misadjustment with "distortion", the situation becomes akin to an operational rate-distortion (R/D) problem in which, for each possible "rate", the goal is to find the model yielding t lie minimum distortion. In practice, however, only a finite palette of models is available, the complexity of which is limited by computational reasons: therefore, the entire data vector often proves too "non-stationary" for any single model. If the process is just slowly drifting, adaptive systems are probably the best choice; on the other hand, a wide class of signals exhibits a series of rather abrupt transition between locally regular portions (e.g. speech, images). In this case a common solution is to partition the data uniformly so that the resulting pieces are small enough to appear stationary with respect to the available models. However, if the goal is again to obtain an overall modeling which is optimal in the above R/D sense, it is necessary that the segmentation be a free variable in the modelization process; this is however not the case if a fixed-size time windowing is used. Note that now the reachable points in the R/D plane are in fact indexed not just by a model but by a segmentation/model-sequence pair; their number therefore grows exponentially with the size of the data vector. This thesis is concerned with the development of efficient techniques to explore this R/D set and to determine its operational lower bound for specific signal processing problems. It will be shown that, under very mild hypotheses, many practical applications dealing with nonstationary data sets can be cast as a R/D optimization problem involving segmentation, which can in turn be solved using polynomial-time dynamic programming techniques. The flexibility of the approach will be demonstrated by a very diverse set of examples: after a general overview of the various facets of the dynamic segmentation problem in Chapter 2, Chapter 3 will use the framework to determine an operational R/D bound for the approximation of piecewise polynomial function with respect to wavelet-based approximation; Chapter 4 will show its relevant to compression problems, and in particular to speech coding based on linear prediction and to arithmetic coding for binary sources; finally, in Chapter 5, an efficient data hiding scheme for PCM audio signals will be described, in which the optimal power allocation for the hidden data is determined with respect to the time-varying characteristics of the host signal.

EPFL1999