"Can you hear me now?" | EPFL Graph Search

This thesis deals with signal-based methods that predict how listeners perceive speech quality in telecommunications. Such tools, called objective quality measures, are of great interest in the telecommunications industry to evaluate how new or deployed systems affect the end-user quality of experience. Two widely used measures, ITU-T Recommendations P.862 âPESQâ and P.863 âPOLQAâ, predict the overall listening quality of a speech signal as it would be rated by an average listener, but do not provide further insight into the composition of that score. This is in contrast to modern telecommunication systems, in which components such as noise reduction or speech coding process speech and non-speech signal parts differently. Therefore, there has been a growing interest for objective measures that assess different quality features of speech signals, allowing for a more nuanced analysis of how these components affect quality. In this context, the present thesis addresses the objective assessment of two quality features: background noise intrusiveness and speech intelligibility. The perception of background noise is investigated with newly collected datasets, including signals that go beyond the traditional telephone bandwidth, as well as Lombard (effortful) speech. We analyze listener scores for noise intrusiveness, and their relation to scores for perceived speech distortion and overall quality. We then propose a novel objective measure of noise intrusiveness that uses a sparse representation of noise as a model of high-level auditory coding. The proposed approach is shown to yield results that highly correlate with listener scores, without requiring training data. With respect to speech intelligibility, we focus on the case where the signal is degraded by strong background noises or very low bit-rate coding. Considering that listeners use prior linguistic knowledge in assessing intelligibility, we propose an objective measure that works at the phoneme level and performs a comparison of phoneme class-conditional probability estimations. The proposed approach is evaluated on a large corpus of recordings from public safety communication systems that use low bit-rate coding, and further extended to the assessment of synthetic speech, showing its applicability to a large range of distortion types. The effectiveness of both measures is evaluated with standardized performance metrics, using corpora that follow established recommendations for subjective listening tests.

Novel speech processing techniques for robust automatic speech recognition

Vivek Tyagi

The goal of this thesis is to develop and design new feature representations that can improve the automatic speech recognition (ASR) performance in clean as well noisy conditions. One of the main shortcomings of the fixed scale (typically 20-30 ms long analysis windows) envelope based feature such as MFCC, is their poor handling of the non-stationarity of the underlying signal. In this thesis, a novel stationarity-synchronous speech spectral analysis technique has been proposed that sequentially detects the largest quasi-stationary segments in the speech signal (typically of variable lengths varying from 20-60 ms), followed by their spectral analysis. In contrast to a fixed scale analysis technique, the proposed technique provides better time and frequency resolution, thus leading to improved ASR performance. Moving a step forward, this thesis then outlines the development of theoretically consistent amplitude modulation and frequency modulation (AM-FM) techniques for a broad band signal such as speech. AM-FM signals have been well defined and studied in the context of communications systems. Borrowing upon these ideas, several researchers have applied AM-FM modeling for speech signals with mixed results. These techniques have varied in their definition and consequently the demodulation methods used therein. In this thesis, we carefully define AM and FM signals in the context of ASR. We show that for a theoretically meaningful estimation of the AM signals, it is important to constrain the companion FM signal to be narrow-band. Due to the Hilbert relationships, the AM signal induces a component in the FM signal which is fully determinable from the AM signal and hence forms the redundant information. We present a novel homomorphic filtering technique to extract the leftover FM signal after suppressing the redundant part of the FM signal. The estimated AM message signals are then down-sampled and their lower DCT coefficients are retained as speech features. We show that this representation is, in fact, the exact dual of the real cepstrum and hence, is referred to as fepstrum. While Fepstrum provides amplitude modulations (AM) occurring within a single frame size of 100ms, the MFCC feature provides static energy in the Mel-bands of each frame and its variation across several frames (the deltas). Together these two features complement each other and the ASR experiments (hidden Markov model and Gaussian mixture model (HMM-GMM) based) indicate that Fepstrum feature in conjunction with MFCC feature achieve significant ASR improvement when evaluated over several speech databases. The second half of this thesis deals with the noise robust feature extraction techniques. We have designed an adaptive least squares filter (LeSF) that enhances a speech signal corrupted by broad band noise that can be non-stationary. This technique exploits the fact that the autocorrelation coefficients of a broad-band noise decay much more rapidly with increasing time lag as compared to those of the speech signal. This is especially true for voiced speech as it consists of several sinusoids at the multiples of the fundamental frequency. Hence the autocorrelation coefficients of the voiced speech are themselves periodic with period equal to the pitch period. On the other hand, the autocorrelation coefficients of a broad band noise are rapidly decaying with increasing time lag. Therefore, a high order (typically 100 tap) least square filter that has been designed to predict a noisy speech signal (speech + additive broad band noise) will predict more of the clean speech components than the broad band noise. This has been analytically proved in this thesis and we have derived analytic expressions for the noise rejection achieved by such a least squares filter. This enhancement technique has led to significant ASR accuracy in the presence of real life noises such as factory noise and aircraft cockpit noise. Finally, the last two chapters of this thesis deal with feature level noise robustness technique. Unlike the least squares filtering that enhances the speech signal itself (in the time domain), the feature level noise robustness techniques as such do not enhance the speech signal but rather boosts the noise-robustness of the speech features that usually are non-linear functions of the speech signal's power spectrum. The techniques investigated in this thesis provided a significant improvement in the ASR performance for the clean as well noisy acoustic conditions.

EPFL2006

Sparsity-based Source Separation with Scale-Discretized Steerable Wavelets

Pierre Vandergheynst, Yves Wiaux

In this talk, we discuss the issue of source separation in a signal on the sphere or on the plane, relying on the sparsity of one signal component in a scale-discretized steerable wavelet basis. The steerability of wavelets allows to probe in detail the local morphology of a signal at each analysis scale. It gives access to local measures of signed-intensity, orientation, elongation, etc. The scale discretization of the wavelets allows the reconstruction of the signal analyzed. In this context, local directional features can be identified at any scale from their sparsity in wavelet space, and reconstructed after their separation from other signal components. In cosmology, this approach reveals to be of great interest for the search of topological defects in the cosmic microwave background signal, such as cosmic strings.

2008

Grapheme-based Automatic Speech Recognition using Probabilistic Lexical Modeling

Ramya Rasipuram

Automatic speech recognition (ASR) systems incorporate expert knowledge of language or the linguistic expertise through the use of phone pronunciation lexicon (or dictionary) where each word is associated with a sequence of phones. The creation of phone pronunciation lexicon for a new language or domain is costly as it requires linguistic expertise, and includes time and money. In this thesis, we focus on effective building of ASR systems in the absence of linguistic expertise for a new domain or language. Particularly, we consider graphemes as alternate subword units for speech recognition. In a grapheme lexicon, pronunciation of a word is derived from its orthography. However, modeling graphemes for speech recognition is a challenging task for two reasons. Firstly, grapheme-to-phoneme (G2P) relationship can be ambiguous as languages continue to evolve after their spelling has been standardized. Secondly, as elucidated in this thesis, typically ASR systems directly model the relationship between graphemes and acoustic features; and the acoustic features depict the envelope of speech, which is related to phones. In this thesis, a grapheme-based ASR approach is proposed where the modeling of the relationship between graphemes and acoustic features is factored through a latent variable into two models, namely, acoustic model and lexical model. In the acoustic model the relationship between latent variables and acoustic features is modeled, while in the lexical model a probabilistic relationship between latent variables and graphemes is modeled. We refer to the proposed approach as probabilistic lexical modeling based ASR. In the thesis we show that the latent variables can be phones or multilingual phones or clustered context-dependent subword units; and an acoustic model can be trained on domain-independent or language-independent resources. The lexical model is trained on transcribed speech data from the target domain or language. In doing so, the parameters of the lexical model capture a probabilistic relationship between graphemes and phones. In the proposed grapheme-based ASR approach, lexicon learning is implicitly integrated as a phase in ASR system training as opposed to the conventional approach where first phone pronunciation lexicon is developed and then a phone-based ASR system is trained. The potential and the efficacy of the proposed approach is demonstrated through experiments and comparisons with other standard approaches on ASR for resource rich languages, nonnative and accented speech, under-resourced languages, and minority languages. The studies revealed that the proposed framework is particularly suitable when the task is challenged by the lack of both linguistic expertise and transcribed data. Furthermore, our investigations also showed that standard ASR approaches in which the lexical model is deterministic are more suitable for phones than graphemes, while probabilistic lexical model based ASR approach is suitable for both. Finally, we show that the captured grapheme-to-phoneme relationship can be exploited to perform acoustic data-driven G2P conversion.