Qiguang Lin

Robust hands‐free speech recognition

When speech recognition technology moves from the laboratory to real‐world applications, there is increasing need for robustness. This paper describes a system of microphone arrays and neural networks (MANN) for robust hands‐free speech recognition. MANN has the advantage that existing speech recognition systems can directly be deployed in practical adverse environments where distant‐talking sound pickup is required. No retraining nor modification of the recognizers is necessary. MANN consists of two synergistic components: (1) signal enhancement by microphone arrays and (2) feature adaptation by neural network computing. High‐quality sound capture by the microphone array enables successful feature adaptation by the neural network to mitigate environmental interference. Through neural network computation, a matched training and testing condition is approximated which typically elevates performance of speech recognition. Both computer‐simulated and real‐room speech input are used to evaluate the capability...

Vocal tract simulations based on fluid dynamic analysis

An alternate approach to speech synthesis based on direct numerical solution of Navier–Stokes (NS) and Reynolds‐averaged‐Navier–Stokes (RANS) equations is described. Unlike the traditional methods based on linear acoustic theory, the NS and RANS formulations are not limited by the assumptions of linearity, negligible viscous effects, and plane‐wave propagation. The expected results are high‐quality synthesis and a new parametrization of speech for applications in automatic recognition and low bit‐rate coding resulting from a parsimonious modeling of articulatory shapes and dynamics. In the present formulation, the Navier–Stokes equations are discretized and solved using a finite‐difference method. An initial application involves a 2‐D simulation of flow through ideal channels (straight, dilating, and constricting tubes) with rigid walls and constant boundary conditions (constant flow velocity at inlet, zero pressure at outlet). As expected for these simple geometries, the resonance frequencies correspond ...

Open‐loop acoustic‐to‐articulatory mapping

A new design is presented for open‐loop estimation of articulatory parameters from an acoustic speech signal. Previous systems for acoustic‐to‐articulatory mapping generally involve an optimization loop which adapts synthetic speech to an arbitrary speech input. The open‐loop method (i.e., without any optimization), aims to provide an accurate estimation of articulatory parameters at a low computational cost. The technique is used as a component of an adaptive voice mimic system. Model shapes corresponding to the natural signal are found by searching a precomputed table. The table associates vocal tract shapes to their corresponding spectra obtained by linear acoustic simulation. A new metric based on poles from linear predictive analysis is proposed to compare the natural spectrum to the precomputed synthetic spectra. Nearly real‐time processing is achieved on a workstation by introducing a two‐step search strategy. The resulting representation is known to provide an efficient parametrization of the spee...

Acoustics and synthesis of nasalization

In this paper acoustics and synthesis of nasalization is studied based on a comprehensive computer model of the vocal tract: TRACTTALK. TRACTTALK simulates the vocal‐tract system in the frequency domain and derives the time‐domain equivalent to produce sound output. It incorporates all important components of the system and decomposes the transfer function into its zero and pole parts. Such a decomposition enables one to accurately estimate poles and zeros of a nasalized sound. First, temporal trajectories of poles/zeros are examined as a function of the velopharyngeal opening, the presence of the nasal sinuses, and other articulatory parameters. This is in follow‐up to previous work [Flanagan, AT&T Bell Labs. internal report (1983)]. The attempt is to systematically characterize the pole/zero pattern of nasalization for improving the performance of formant tracking and feature labeling algorithms. Secondly, synthesis of nasalization is described using TRACTTALK. Listening experiments are conducted to ass...

Use of TRACTTALK for adaptive voice mimic

Various speech‐processing technologies necessitate parametrization of the speech waveform. Cepstrum coefficients (including their derivatives and variants) are to date commonly used in speech and speaker recognition. This paper seeks more compact parametric description of speech information based on the adaptive voice mimic [Flanagan et al., 780–791 (1980)]. The mimic system utilizes an articulatory‐based speech synthesizer to generate synthetic speech, which is adapted to arbitrary speech input. The perceptually weighted spectral difference between the input and synthesized speech is next minimized by optimizing the underlying articulatory parameters until the difference is driven below a predetermined level. The resultant representation, adapted moment by moment, provides efficient parametrization of the signal information by which the problems of speech synthesis, speech recognition, and low bit‐rate speech coding are coalesced into a compact framework. In this paper, an articulatory speech synthesizer...

Vowel recognition using an articulatory representation

An alternate approach to speech recognition based on an articulatory representation of speech is proposed. Unlike traditional methods based on Fourier or cepstrum representations, the articulatory description of speech provides a compact parametrization linked to physiological properties. The expected results are robust speaker‐independent speech recognition. In fact, physiologically related parameters should be useful for handling variability across speakers. In this work, a three‐parameter articulatory representation is applied to vowel recognition. These parameters (location and size of the main constriction in the vocal tract and the ratio of lip length to mouth aperture) are estimated using a codebook search strategy. Before the recognition process, a pattern database is built using a corpus of ten repetitions of nine different vowels. Vector quantization is then applied to this corpus to obtain several representative articulatory vectors, or centroids, for each vowel. In the recognition stage, artic...