Patent Application
20070260454
For a better understanding of the present invention, a preferred embodiment, which is intended purely by way of example and is not to be construed as limiting, will now be described with reference to the attached drawings, wherein:
The following discussion is presented to enable a person skilled in the art to make and use the invention. Various modifications to the embodiments will be readily apparent to those skilled in the art, and the generic principles herein may be applied to other embodiments and applications without departing from the spirit and scope of the present invention. Thus, the present invention is not intended to be limited to the embodiments shown, but is to be accorded the widest scope consistent with the principles and features disclosed herein and defined in the attached claims.
The present invention relates to an automatic speech recognition system including a noise reduction system based on the Spectral Attenuation Technique, and in particular on the Ephraim-Malah spectral attenuation rule, wherein the global formula of the gain Gk(γk, ξk) is unchanged, whereas the estimates of the a priori and the a posteriori signal-to-noise ratios {circumflex over (ξ)}k(m), {circumflex over (γ)}k(m) are modified by making them dependent on a noise weighting factor α(m) and on a spectral flooring factor β(m), as follows:
where:
The noise weighting factor α(m) and the spectral flooring factor β(m) are a function of the global signal-to-noise ratio SNR(m), which is defined as:
The values indicated in formulas (12) and (13) are intended purely by way of example and are not to be construed as limiting. In general, other values could be usefully employed, while maintaining the general development of the noise weighting factor α(m) and of the spectral flooring factor β(m) versus the global signal-to-noise ratio SNR(m).
In particular, the noise weighting factor α(m) versus the global signal-to-noise ratio SNR(m) should have a first substantially constant value when the global signal-to-noise ratio SNR(m) is lower than a first threshold, a second substantially constant value lower than the first substantially constant value when the global signal-to-noise ratio SNR(m) is higher than a second threshold, and values decreasing from the first substantially constant value to the second substantially constant value when the global signal-to-noise ratio SNR(m) increases from the first threshold to the second threshold.
The spectral flooring factor β(m) versus the global signal-to-noise ratio SNR(m) should have a first substantially constant value when the global signal-to-noise ratio SNR(m) is lower than a first threshold, a second substantially constant value higher than the first substantially constant value when the global signal-to-noise ratio SNR(m) is higher than a second threshold, and values increasing from the first substantially constant value to the second substantially constant value when the global signal-to-noise ratio SNR(m) increases from the first threshold to the second threshold. The developments may be piece-wise lines, as shown in
The estimate |Dk(m)|2 of the power spectrum of the noisy speech in formulas (9), (10) and (11) is computed by means of a first-order recursion as disclosed in the aforementioned Noise Estimation Techniques for Robust Speech Recognition.
Preferably, the first-order recursion may be implemented in conjunction with a standard energy-based Voice Activity Detector, which is well-known system which detects presence or absence of speech based on a comparison of the total energy of the speech signal with an adaptive threshold and outputs a Boolean flag (VAD) having a “true” value when voice is present and a “false” value when voice is absent. When a standard energy-based Voice Activity Detector is used, the estimate |{circumflex over (D)}k(m)|2 of the power spectrum of the noisy speech may be computed as follows:
where λ is a weighting factor which controls the update speed of the recursion and ranges between 0 and 1, preferably has a value of 0.9, μ is a multiplication factor which controls the allowed dynamics of the noise and preferably has a value of 4.0, and σ(m) is the noise standard deviation, estimated as follows:
σ2(m)=λσ2(m−1)+(1−λ(|Yk(m)|2−|{circumflex over (D)}k(m)|2)2 (15)
A large experimental work has been performed to validate the invention, and some results, which may be useful to highlight the features of the invention, are hereinafter reported.
In particular, experiments were conducted with an automatic speech recognition system, using noise reduction with the standard Ephraim-Malah spectral attenuation and with the noise reduction proposed in the invention. The automatic speech recognition system has been trained for the target languages using large, domain and task independent corpora, not collected in noisy environments and without added noise.
The experiment was performed on the Aurora3 corpus, that is a standard corpus defined by the ETSI Aurora Project for noise reduction tests, and made of connected digits recorded in car in several languages (Italian, Spanish and German). An high mismatch test set and a noisy component of the training set (used as test set) were employed.
The modification of the Ephraim-Malah spectral attenuation rule according to the invention produces an average error reduction of 28.9% with respect to the state of the art Wiener Spectral Subtraction, and an average error reduction of 22.9% with respect to the standard Ephraim-Malah Spectral Attenuation Rule. The average error reduction with respect to no de-noising is 50.2%.
Finally, it is clear that numerous modifications and variants can be made to the present invention, all falling within the scope of the invention, as defined in the appended claims.
| Number | Date | Country | |
|---|---|---|---|
| Parent | PCT/EP04/50816 | May 2004 | US |
| Child | 11598705 | US |
