If you want to have the list of publications issued from a specific Individual Project (IP), write in the search field (IM2.IP). IP can have the following value: DMA, AP, VP, MPR, MCA, HMI, ISD, BMI
If you want to find joint publications between IPs, write in the search field (joint), click on search and then click on Keywords
If you want to display all the publications for a specific author, use the shortcut called -Authors- located in the main menu
Abstract: We present a proposal of a kernel-based model for large vocabulary continuous speech recognizer. The continuous speech recognition is described as a problem of finding the best phoneme sequence and its best time span, where the phonemes are generated from all permissible word sequences. A non-probabilistic score is assigned to every phoneme sequence and time span sequence, according to a kernel-based acoustic model and a kernel-based language model. The acoustic model is described in terms of segments, where each segment corresponds to a whole phoneme, and it generalizes Segmental Models for the non-probabilistic setup. The language model is based on discriminative language model recently proposed by Roark et al. (2007). We devise a loss function based on the word error rate and present a large margin training procedure for the kernel models, which aims at minimizing this loss function. Finally, we discuss the practical issues of the implementation of kernel-based continuous speech recognition model by presenting an efficient iterative algorithm and considering the decoding process. We conclude the chapter by a brief discussion on the model limitations and future work. This chapter does not introduce any experimental results.
Abstract: A speech/audio codec based on Frequency Domain Linear Prediction (FDLP) exploits auto-regressive modeling to approximate instantaneous energy in critical frequency sub-bands of relatively long input segments. The current version of the FDLP codec operating at 66 kbps has been shown to provide comparable subjective listening quality results to state-of-the-art codecs on similar bit-rates even without employing standard blocks such as entropy coding or simultaneous masking. This paper describes an experimental work to increase compression efficiency of the FDLP codec by employing entropy coding. Unlike conventional Huffman coding employed in current speech/audio coding systems, we describe an efficient way to exploit arithmetic coding to entropy compress quantized spectral magnitudes of the sub-band FDLP residuals. Such an approach provides 11\% ( 3 kbps) bit-rate reduction compared to the Huffman coding algorithm ( 1 kbps).
Abstract: We describe novel audio coding technique designed to be utilized at medium bit-rates. Unlike classical state-of-the-art audio coders that are based on short-term spectra, our approach uses relatively long temporal segments of audio signal in critical-band-sized sub-bands. We apply auto-regressive model to approximate Hilbert envelopes in frequency sub-bands. Residual signals (Hilbert carriers) are demodulated and thresholding functions are applied in spectral domain. The Hilbert envelopes and carriers are quantized and transmitted to the decoder. Our experiments focused on designing audio coder to provide broadcast radio-like quality audio around $10-20$kbps. Objective quality measures indicate comparable performance with the 3GPP-AMR speech codec standard for both speech and non-speech signals.
Abstract: Using phone posterior probabilities has been increasingly explored for improving automatic speech recognition (ASR) systems. In this paper, we propose two approaches for hierarchically enhancing these phone posteriors, by integrating long acoustic context, as well as prior phonetic and lexical knowledge. In the first approach, phone posteriors estimated with a Multi-Layer Perceptron (MLP), are used as emission probabilities in HMM forward-backward recursions. This yields new enhanced posterior estimates integrating HMM topological constraints (encoding specific phonetic and lexical knowledge), and context. posteriors are post-processed by a secondary MLP, in order to learn inter and intra dependencies between the phone posteriors. These dependencies are prior phonetic knowledge. The learned knowledge is integrated in the posterior estimation during the inference (forward pass) of the second MLP, resulting in enhanced phone posteriors. We investigate the use of the enhanced posteriors in hybrid HMM/ANN and Tandem configurations. We propose using the enhanced posteriors as replacement, or as complementary evidences to the regular MLP posteriors. The proposed method has been tested on different small and large vocabulary databases, always resulting in consistent improvements in frame, phone and word recognition rates.
Abstract: The use of local phoneme posterior probabilities has been increasingly explored for improving speech recognition systems. Hybrid hidden Markov model / artificial neural network (HMM/ANN) and Tandem are the most successful examples of such systems. In this thesis, we present a principled framework for enhancing the estimation of local posteriors, by integrating phonetic and lexical knowledge, as well as long contextual information. This framework allows for hierarchical estimation, integration and use of local posteriors from the phoneme up to the word level. We propose two approaches for enhancing the posteriors. In the first approach, phoneme posteriors estimated with an ANN (particularly multi-layer Perceptron - MLP) are used as emission probabilities in HMM forward-backward recursions. This yields new enhanced posterior estimates integrating HMM topological constraints (encoding specific phonetic and lexical knowledge), and long context. In the second approach, a temporal context of the regular MLP posteriors is post-processed by a secondary MLP, in order to learn inter and intra dependencies among the phoneme posteriors. The learned knowledge is integrated in the posterior estimation during the inference (forward pass) of the second MLP, resulting in enhanced posteriors. The use of resulting local enhanced posteriors is investigated in a wide range of posterior based speech recognition systems (e.g. Tandem and hybrid HMM/ANN), as a replacement or in combination with the regular MLP posteriors. The enhanced posteriors consistently outperform the regular posteriors in different applications over small and large vocabulary databases.
Abstract: Frequency Domain Linear Prediction (FDLP) represents a technique for auto-regressive modelling of Hilbert envelopes of a signal. In this paper, we propose a speechcoding technique that uses FDLP in Quadrature Mirror Filter (QMF) sub-bands of short segments of the speech signal (25 ms). Line Spectral Frequency parameters related to au-toregressive models and the spectral components of the residual signals are transmitted. For simulating the effects of lossy transmission channels, bit-packets are dropped randomly. In the objective and subjective quality evaluations, the proposed FDLP speech codec is judged to be more resilient to bit-packet losses compared to the state-of-the-art Adaptive Multi-Rate Wide-Band (AMR-WB) codec at 12 kbps.
Abstract: Frequency Domain Linear Prediction (FDLP) represents a technique for auto-regressive modelling of Hilbert envelopes of a signal. In this paper, we propose a speechcoding technique that uses FDLP in Quadrature Mirror Filter (QMF) sub-bands of short segments of the speech signal (25 ms). Line Spectral Frequency parameters related to au-toregressive models and the spectral components of the residual signals are transmitted. For simulating the effects of lossy transmission channels, bit-packets are dropped randomly. In the objective and subjective quality evaluations, the proposed FDLP speech codec is judged to be more resilient to bit-packet losses compared to the state-of-the-art Adaptive Multi-Rate Wide-Band (AMR-WB) codec at 12 kbps.
Abstract: beginabstract We present a new filter bank design method for subband adaptive beamforming. Filter bank design for adaptive filtering poses many problems not encountered in more traditional applications such as subband coding of speech or music. The popular class of perfect reconstruction filter banks is not well-suited for applications involving adaptive filtering because perfect reconstruction is achieved through alias cancellation, which functions correctly only if the outputs of individual subbands are emphnot subject to arbitrary magnitude scaling and phase shifts. In this work, we design analysis and synthesis prototypes for modulated filter banks so as to minimize each aliasing term individually. We then show that the emphtotal response error can be driven to zero by constraining the analysis and synthesis prototypes to be emphNyquist($M$) filters. We show that the proposed filter banks are more robust for aliasing caused by adaptive beamforming than conventional methods. Furthermore, we demonstrate the effectiveness of our design technique through a set of automatic speech recognition experiments on the multi-channel, far-field speech data from the emphPASCAL Speech Separation Challenge. In our system, speech signals are first transformed into the subband domain with the proposed filter banks, and thereafter the subband components are processed with a beamforming algorithm. Following beamforming, post-filtering and binary masking are performed to further enhance the speech by removing residual noise and undesired speech. The experimental results prove that our beamforming system with the proposed filter banks achieves the best recognition performance, a 39.6\% word error rate (WER), with half the amount of computation of that of the conventional filter banks while the perfect reconstruction filter banks provided a 44.4\% WER. endabstract
Abstract: Unlike classical state-of-the-art coders that are based on short-term spectra, our approach uses relatively long temporal segments of audio signal in critical-band-sized sub-bands. We apply auto-regressive model to approximate Hilbert envelopes in frequency sub-bands. Residual signals (Hilbert carriers) are demodulated and thresholding functions are applied in spectral domain. The Hilbert envelopes and carriers are quantized and transmitted to the decoder. Our experiments focused on designing speech/audio coder to provide broadcast radio-like quality audio around 15-25kbps. Obtained objective quality measures, carried out on standard speech recordings, were compared to the state-of-the-art 3GPP-AMR speechcoding system.
Abstract: In this paper we propose a new technique for robust keyword spotting that uses bidirectional Long Short-Term Memory (BLSTM) recurrent neural nets to incorporate contextual information in speech decoding. Our approach overcomes the drawbacks of generative HMM modeling by applying a discriminative learning procedure that non-linearly maps speech features into an abstract vector space. By incorporating the outputs of a BLSTM network into the speech features, it is able to make use of past and future context for phoneme predictions. The robustness of the approach is evaluated on a keyword spotting task using the HUMAINE Sensitive Artificial Listener (SAL) database, which contains accented, spontaneous, and emotionally colored speech. The test is particularly stringent because the system is not trained on the SAL database, but only on the TIMIT corpus of read speech. We show that our method prevails over a discriminative keyword spotter without BLSTM-enhanced feature functions, which in turn has been proven to outperform HMM-based techniques.
Abstract: In this paper we propose an extension of the very low bit-rate speechcoding technique, exploiting predictability of the temporal evolution of spectral envelopes, for wide-band audio coding applications. Temporal envelopes in critically band-sized sub-bands are estimated using frequency domain linear prediction applied on relatively long time segments. The sub-band residual signals, which play an important role in acquiring high quality reconstruction, are processed using a heterodyning-based signal analysis technique. For reconstruction, their optimal parameters are estimated using a closed-loop analysis-by-synthesis technique driven by a perceptual model emulating simultaneous masking properties of the human auditory system. We discuss the advantages of the approach and show some properties on challenging audio recordings. The proposed technique is capable of encoding high quality, variable rate audio signals on bit-rates below $1$bit/sample.
