Block-diagonal covariance joint subspace tying and model compensation for noise robust automatic speech recognition

[0001] This application claims the benefit of U.S. Provisional Application No. 60/659,054, filed on Mar. 4, 2005. The disclosure of the above application is incorporated herein by reference in its entirety for any purpose.

[0002] The present invention generally relates to automatic speech recognition, and relates in particular to noise robust automatic speech recognition.

[0003] Embedded noise robust automatic speech recognition (ASR) systems need to conserve memory due to the small size and limited resources of devices such as cell phones, car navigation, digital TVs, and home appliances. However, ASR systems are notorious for consuming large amounts of computational resources, including Random Access Memory (RAM). This tendency of ASR systems can be especially problematic in embedded devices that also need to allocate such resources for other functions that often need to run concurrently with ASR functions. Yet, reducing the amount of memory consumed by a noise robust ASR heavily impacts recognition accuracy and/or robustness to noise.

[0004] Referring to FIG. 1, model domain methods try to improve the performance of pattern matching by modifying the acoustic models so that they are adapted to the current noise level, while leaving the input signal 100 unchanged. In particular, a noise estimation module 104 estimates noise in the input signal 100, and model compensation module 106 adjusts the acoustic models 108 based on these noise estimates. Then, extracted features obtained from the unmodified input signal 100 by feature extraction module 102 are pattern matched to the adjusted acoustic models 108 by pattern matching module 110 to achieve recognition 112.

[0005] What is needed is a way to reduce the memory requirements of embedded noise robust ASR systems with reduced impact on recognition accuracy and/or robustness to noise. The present invention fulfills this need by making several changes to a noise robustness system employing a model domain method.

[0006] In accordance with the present invention, model compression is combined with model compensation. Model compression is needed in embedded ASR to reduce the size and the computational complexity of compressed models. Model-compensation is used to adapt in real-time to changing noise environments. The present invention allows for the design of smaller ASR engines (memory consumption reduced to up to one-sixth) with reduced impact on recognition accuracy and/or robustness to noises.

[0007] Further areas of applicability of the present invention will become apparent from the detailed description provided hereinafter. It should be understood that the detailed description and specific examples, while indicating the preferred embodiment of the invention, are intended for purposes of illustration only and are not intended to limit the scope of the invention.

[0012] The following description of the preferred embodiments is merely exemplary in nature and is in no way intended to limit the invention, its application, or uses.

[0013] In some embodiments, the present invention combines sub-space tying for model compression with alpha-Jacobian model-compensation for noise robustness to achieve a compact noise robust speech recognition system. Unfortunately this combination cannot be accomplished directly as the subspace tying structure does not allow for model-compensation. This difficulty arises because the distortion function used in model compensation requires a full space transformation (full dimensionality) of the acoustic models that invalidates the tying structure.

[0014] One area of interest in the present invention is the present solution to this issue. Specifically, a model compensation distortion function is designed that does not invalidate the tying structure, thus allowing for the coexistence of subspace tying and model-compensation. The design of the model compensation distortion function can be accomplished by making several changes in a noise robust ASR system to the following modules: (a) front-end analysis: the front-end whitening matrix can be block-diagonal to isolate a set independent subspaces (block-diagonal covariance matrix); (b) model-compensation: the model-compensation distortion function can be operating independently on the same subspaces identified by the front-end analysis (and cannot be a full-space transformation); and (c) subspace model compression: the subspaces used for the tying can be aligned with the independent subspaces defined in the front-end.

[0015] One ingredient of this method can be in the definition of the subspaces corresponding to the block-diagonal whitening matrix in the front-end. These subspaces need to be large enough to allow a good coverage of the speech signal correlation structure in the front-end and in the model-compensation step, but small enough to allow a low distortion error from the subspace tying step. In general, the subspace definition is an NP-Complete problem for which there is no computable exact solution, but for which an interactive converging algorithm can be provided.

[0016] The whitening matrix or matrices can take various forms depending on the characteristics of the independent subspaces. For example, in some embodiments, the independent subspaces can span over different time frames, and the whitening matrices include decorrelation across a 2-dimensional time-frequency axis. Also, in additional or alternative embodiments, such 2-D decorrelation matrices are decomposable as discrete cosine transform in the frequency domain and time derivative in the time domain.

[0017] Turning to FIG. 2, real-time implementation of a model compensation distortion function in accordance with the present invention includes a number of different components. Such components can include speech input 250, noise estimation 252, power spectral energy estimation by Short Time Fourier Transform (STFT) or Wavelets decomposition 254, band 256, log compression or power-law compression function 258, and tying topology (subspace definition structure) 260. All components in the three main blocks of front-end analysis 200, model-compensation 202, and subspace Gaussian distribution computation 204 are split and aligned to follow the subspace definition structure. The decorrelation matrices 206 operate independently on blocks 208 of log filter-bank energies. This architecture allows for the model-compensation to work effectively on each subspace without affecting the subspace tying structure. This capability allows for efficient model-compensation of subspace compressed acoustic models, which in turn allows a considerable reduction in system size and a considerable improvement in speed.

[0018] The model compensation distortion technique of the present invention allows reduction of the acoustic models size by up to ⅙^th of the initial size, and reduces the computational load to up to ⅓^rd of the initial while allowing great robustness to noise thanks to the usage of model-compensation. The complexity of the model compensation is also reduced because of the smaller set of distributions to compensate.

[0019] Turning now to FIG. 3, results illustrate performance of the noise robust ASR system according to the present invention for a Car-Navigation task in noisy conditions. Jacobian model compensation is applied in all three cases. The “Sub-space” case 300 shows the performance of the proposed invention. The previous embedded model compression method allowing for model compensation, the “Full-space” case 302, does not provide a good recognition rate, mainly because full-space compression introduces too much distortion in the acoustic models. The proposed method provides better performance, very close to uncompressed models, the “untied” case 304, but with smaller size and a faster real-time factor.

[0020] It is envisioned that a similar approach can be employed for speaker adaptation, with subspace transformations (such as MLLR constrained to subspaces). For example, subspace tied acoustic model whitening can be employed with model compensation and an additional subspace tying regarding compensated acoustic models for update purposes (store to RAM or flash ROM, etc.).

[0021] The description of the invention is merely exemplary in nature and, thus, variations that do not depart from the gist of the invention are intended to be within the scope of the invention. Such variations are not to be regarded as a departure from the spirit and scope of the invention.