Scalable distributed speech recognition using multi-frame GMM-based block quantization

This paper describes a noise-robust Distributed Speech Recognition (DSR) front-end using a combination of conventional Mel-cepstral Coefficient (MFCC) and Line Spectral Frequencies (LSF). These features are adequately transformed and reduced in a multi-stream scheme using Karhunen-Loeve Transform (KLT). We investigate the performance of a new front-end DSR in terms of recognition accuracy in adverse conditions as well as in terms of dimensionality reduction. Our results showed that for highly noisy speech, the proposed transformation scheme leads to a significant improvement in recognition accuracy on Aurora 2 task.

In this paper, a new noise front-end is proposed to improve the performance of Distributed Speech Recognition (DSR) system using a combination of conventional Mel-Cepstral Coefficients (MFCC) and Mel-Line Spectral Frequencies (MLSF). These features are adequately transformed and reduced in a multi-stream scheme using Karhunen-Loeve Transform (KLT). We investigate the performance of a new front-end DSR in terms of recognition accuracy in adverse conditions as well as in terms of dimensionality reduction. Our results showed that for highly noisy speech, using the proposed transformation scheme MLSF-KLT leads to a significant improvement in recognition accuracy on Aurora 2 task.

We propose a novel exemplar-based feature enhancement method for automatic speech recognition which uses coupled dictionaries: an input dictionary containing atoms sampled in the modulation (envelope) spectrogram domain and an output dictionary with atoms in the Mel or full-resolution frequency domain. The input modulation representation is chosen for its separation properties of speech and noise and for its relation with human auditory processing. The output representation is one which can be processed by the ASR back-end. The proposed method was investigated on the AURORA-2 and AURORA-4 databases and improved word error rates (WER) were obtained when compared to the system which uses Mel features in the input exemplars. The paper also proposes a hybrid system which combines the baseline and the proposed algorithm on the AURORA-2 database which in turn also yielded improvement over both the algorithms.

This work deals with the packet loss problem in a Distributed Speech Recognition architecture. A packet loss simulation model is first proposed in order to simulate different channel degradation conditions. In these conditions, the performance of our continuous French speech recognition system is evaluated for packets containing different numbers of speech feature vectors. Several reconstruction strategies, to recover lost information, are proposed and evaluated. The results first confirm the intuitive fact that the word error rate obviously increases with the size of the lost packets and with the channel degradation level. However, it is shown that simple reconstruction strategies allow to recover acceptable performance. The most efficient ones are those using interleaving technique to distribute the speech information among packets, combined with interpolation methods to estimate lost acoustic features.

Distributed Speech Recognition (DSR) systems rely on efficient transmission of speech information from distributed clients to a centralized server. Wireless or network communication channels within DSR systems are typically noisy and bursty. Thus, DSR systems must utilize efficient Error Recovery (ER) schemes during transmission of speech information. Some ER strategies, referred to as forward error control (FEC), aim to create redundancy in the source coded bitstream to overcome the effect of channel errors, while others are designed to create spread or delay in the feature stream in order to overcome the effect of bursty channel errors. Furthermore, ER strategies may be designed as a combination of the previously described techniques. This chapter presents an array of error recovery techniques for remote speech recognition applications. This chapter is organized as follows. First, channel characterization and modeling are discussed. Next, media-specific FEC is presented for packet erasure applications, followed by a discussion on media-independent FEC techniques for bit error applications, including general linear block codes, cyclic codes, and convolutional codes. The application of unequal error protection (UEP) strategies utilizing combinations of the aforementioned FEC methods is also presented. Finally, framebased interleaving is discussed as an alternative to overcoming the effect of bursty channel erasures. The chapter concludes with examples of modern standards for channel coding strategies for distributed speech recognition (DSR). 8.1. Distributed Speech Recognition Systems Throughout this chapter various error recovery and detection techniques are discussed. It is therefore necessary to present an overview of a complete experimental DSR system, including feature extraction, a noisy channel model, and an automatic speech recognition engine at the server end (Fig. 8.1). 2 Bengt J. Borgstrom, Alexis Bernard, and Abeer Alwan Feature Extraction Source Coding Channel Coding Channel Decoding Source Decoding Speech Recognition Client Server Noisy Channel input speech

Quantization-based watermarking schemes comprise a class of watermarking schemes that achieves the channel capacity in terms of additive noise attacks. The existence of good high dimensional lattices that can be efficiently implemented and incorporated into watermarking structures, made quantization-based watermarking schemes of practical interest. Because of the structure of the lattices, watermarking schemes making use of them are vulnerable to non-additive operations, like amplitude scaling in combination with additive noise. In this paper, we propose a secure Maximum Likelihood (ML) estimation technique for amplitude scaling factors using subtractive dither. The dither has mainly security purposes and is assumed to be known to the watermark encoder and decoder. We derive the probability density function (PDF) models of the watermarked and attacked data in the presence of subtractive dither. The derivation of these models follows the lines of reference 5, where we derived the PDF models in the absence of dither. We derive conditions for the dither sequence statistics such that a given security level is achieved using the error probability of the watermarking system as objective function. Based on these conditions we are able to make approximations to the PDF models that are used in the ML estimation procedure. Finally, experiments are performed with real audio and speech signals showing the good performance of the proposed estimation technique under realistic conditions.

In distributed speech recognition, vector quantization is used to reduce the number of bits for coding speech features at the user end in order to save energy for transmitting speech feature streams to remote recognizers and reduce data traffic congestion. We notice that the overall bit rate of the transmitted feature streams could be further reduced by not sending redundant frames that can be interpolated at the remote server from received frames. Interpolation introduces errors and may degrade speech recognition. This paper investigates the methods of selecting frames for transmission and the effect of interpolation on recognition. Experiments on a large vocabulary recognizer show that with spline interpolation, the overall frame rate for transmission can be reduced by about 50% with a relative increase in word error rate less than 5.2% for clean and noisy speech.

We have proposed the triple Rayleigh probability density function (TR-PDF) model as the in vivo QUS method of liver fibrosis. Additionally, the double Nakagami probability density function (DN-PDF) model was proposed to evaluate the amount and distribution condition of fat droplets in the fatty-liver livers. The DN-PDF permitted to detect early and moderate fatty livers in previous ex vivo studies with over 15 MHz ultrasound. In our recent studies, we are examining clinical data to determine whether two multi-PDF models can evaluate each disease even at low frequencies. In this talk, the results of in vivo studies for liver fibrosis and fatty liver will be introduced.We have proposed the triple Rayleigh probability density function (TR-PDF) model as the in vivo QUS method of liver fibrosis. Additionally, the double Nakagami probability density function (DN-PDF) model was proposed to evaluate the amount and distribution condition of fat droplets in the fatty-liver livers. The DN-PDF permitted to detect early and moderate fatty livers in previous ex vivo studies with over 15 MHz ultrasound. In our recent studies, we are examining clinical data to determine whether two multi-PDF models can evaluate each disease even at low frequencies. In this talk, the results of in vivo studies for liver fibrosis and fatty liver will be introduced.

This study presents a detailed analysis of probability density functions (PDFs) for aperture-averaged fluctuations of laser radiation power under various atmospheric turbulence conditions. Using numerical simulations, we examined an extensive range of Rytov parameter values, representing different turbulence strengths, and analyzed propagation across all primary diffraction regimes, including plane wave, narrow beam, and spherical wave approximations. Four key PDF models-log-normal, gamma, fractional exponential, and fractional gamma-were evaluated for their ability to model intensity fluctuations, with the optimal PDF identified for various propagation scenarios. The results highlight a continuous transition between PDF models as the scintillation regime changes, offering guidelines to selecting the best-fit PDF under specific propagation conditions. Comparison of simulation results with experimental data confirms that the selected PDFs accurately represent atmospheric turbulence effects on laser beam propagation across various aperture sizes and distances.

Transported probability density function modelling of the vapour phase of an n-heptane jet at diesel engine conditions

Multi-microphone speech recognition integrating beamforming, robust feature extraction, and advanced DNN/RNN backend

PDF models and synthetic model for the wind speed fluctuations based on the resolution of Langevin equation

Automatic speech recognition (ASR) aims at understanding naturally spoken human speech to be used as text inputs to machines. In multi-speaker environments, where multiple speakers are talking simultaneously with a large amount of overlap, a significant performance degradation may occur with conventional ASR systems if they are trained by recordings of single talkers. This paper proposes a multi-speaker ASR method that incorporates speaker embedding information as an additional input. The embedding information for each of the speakers in the training set was extracted as numeric vectors, and all of the embedding vectors were stacked to construct a total speaker profile matrix. The speaker profile matrix from the training dataset enables finding embedding vectors that are close to the speakers of the input recordings in the test conditions, and it helps to recognize the individual speakers’ voices mixed in the input. Furthermore, the proposed method efficiently reuses the decoder from the existing speaker-independent ASR model, eliminating the need for retraining the entire system. Various speaker embedding methods such as i-vector, d-vector, and x-vector were adopted, and the experimental results show 0.33% and 0.95% absolute (3.9% and 11.5% relative) improvements without and with the speaker profile in the word error rate (WER).

We show that the error signal incurred by block quantizing stationary data is nonstationary under some conditions. Examples are presented that indicate that the mean square error is largest at the block edges. When block quantizers are used to encode pictures at low bit rates, this effect tends to make the block edges visible in the reconstructed picture.

The present study aims at the investigation of the effects of turbulence-chemistry interaction on combustion instabilities using a probability density function (PDF) method. The instantaneous quantities in the flow field were decomposed into the Favre-averaged variables and the stochastic fluctuations, which were calculated by unsteady Reynolds averaged Navier-Stokes (U-RANS) equations and the PDF model, respectively. A joint fluctuating velocityfrequency-composition PDF was used. The governing equations are solved by a consistent hybrid finite volume/Monte-Carlo algorithm on triangular unstructured meshes. A nonreacting flow behind a triangular-shaped bluff body flame stabilizer in a rectilinear combustor was simulated by the present method. The results demonstrate the capability of the present method to capture the large-scale coherent structures. The triple decomposition was performed, by dividing the coherent Favre-averaged velocity into time-averaged value and periodical coherent part, to analyze the coherent and incoherent contributions to Reynolds stresses. A simple modification to the coefficients in the turbulent frequency model will help to improve the simulation results. Unsteady flow fields were depicted by streamlines and vorticity contours. Moreover, the association between turbulence production and vorticity saddle points is illustrated.