Obviousness

An analysis of the obviousness of US Patent 10,783,899 ("the '899 patent") under 35 U.S.C. § 103 is provided below. This analysis is based on prior art available before the patent's priority date of February 5, 2016.

Definition of a Person Having Ordinary Skill in the Art (PHOSITA)

A Person Having Ordinary Skill in the Art (PHOSITA) for the '899 patent would be an individual with a Bachelor's or Master's degree in Electrical Engineering, Computer Science, or a related field, and 2-3 years of professional or academic experience in digital signal processing, specifically in the area of speech enhancement and noise reduction. This experience would include familiarity with standard techniques like spectral subtraction, Wiener filtering, voice activity detection (VAD), and the statistical modeling of speech and noise signals.

Analysis of Independent Claims

The '899 patent's core novelty lies in the specific combination of three main concepts:

1. A soft speech detector that outputs a likelihood or probability of speech presence, rather than a binary decision.
2. A dynamic noise overestimation factor that controls the aggressiveness of the noise suppression.
3. The use of the speech likelihood from the soft detector to directly and dynamically control the noise overestimation factor, increasing it during speech pauses and decreasing it during speech activity.

While each element existed in the prior art, their specific combination to solve the problem of babble noise suppression forms the basis of this analysis.

Prior Art Combination Rendering Claims Obvious

A compelling case for obviousness can be made by combining the teachings of:

• Reference 1 (Ephraim & Malah): U.S. Patent 4,811,404, titled "Speech enhancement system" (filed Jan. 28, 1987), which introduces a widely-used speech enhancement algorithm based on Minimum Mean Square Error Short-Time Spectral Amplitude (MMSE-STSA) estimation. A key aspect of this work is the use of an a priori Signal-to-Noise Ratio (SNR) that is dependent on the probability of speech presence.
• Reference 2 (Cohen): "Noise spectrum estimation in adverse environments: Improved minima controlled recursive averaging" (IEEE Transactions on Speech and Audio Processing, Sep. 2003). This paper by Israel Cohen describes an improved method for noise estimation that is crucial for noise suppression algorithms. It acknowledges the problem of over-subtracting noise, which can cause "musical noise," and implicitly motivates controlling the aggressiveness of suppression.
• Reference 3 (Soon, et al.): "A new voice activity detector for very low signal-to-noise ratios" (IEEE International Symposium on Circuits and Systems, 1999). This reference, among many others, teaches the concept of a "soft" Voice Activity Detector (VAD) that outputs a continuous likelihood value for speech presence, rather than a hard binary decision.

Argument for Obviousness of Claim 1 (System Claim)

Claim 1 details a system with a "soft speech detector" and a "noise suppressor" that uses a "dynamic noise overestimation factor" controlled by the detector's output.

1. Soft Speech Detector: Soon et al. explicitly teaches the use of a soft VAD that provides a likelihood of speech presence. A PHOSITA would understand that a probabilistic output is more nuanced than a binary one and allows for finer control of downstream processes like a noise filter.

2. Noise Suppressor with Spectral Weighting: Ephraim & Malah is a foundational reference for noise suppression using spectral weighting (gain functions). The gain function in Ephraim & Malah, which determines how much to attenuate each frequency component, is directly dependent on the a priori SNR, which is itself modulated by the probability of speech presence. This establishes a clear link between a probabilistic speech measure and the calculation of spectral weighting coefficients.

3. Dynamic Noise Overestimation Factor: The '899 patent uses a "dynamic noise overestimation factor" (βoe) to control the aggressiveness of a Wiener filter (see '899 patent, Eq. 9, 11). This is a specific implementation of a more general concept: controlling the degree of noise subtraction. Ephraim & Malah's gain function already performs this function; a lower a priori SNR (indicating a higher probability of speech absence) results in more aggressive suppression. A PHOSITA would recognize that multiplying the estimated noise spectrum by a factor greater than one is a straightforward way to achieve more aggressive noise reduction, a technique known as over-subtraction. Cohen's work on noise estimation highlights the trade-offs involved, motivating a PHOSITA to find a better way to control this over-subtraction.

4. Motivation to Combine: A PHOSITA working on noise suppression in 2015 would be acutely aware of the central challenge: suppressing noise effectively without distorting the desired speech. The problem is particularly difficult with non-stationary noise like babble.

   • The PHOSITA would know from foundational works like Ephraim & Malah that the optimal level of suppression depends on whether speech is present.
   • They would also know from works like Soon et al. that a "soft" VAD provides a more robust and flexible signal for this purpose than a traditional "hard" VAD.
   • Finally, they would be familiar with the common technique of noise over-subtraction (as discussed in the context of Cohen's work) to reduce residual noise during speech pauses, but would also be aware of the risk of distortion if applied during speech.

   Therefore, it would have been obvious to a PHOSITA to combine these concepts. The most direct and predictable way to improve upon the system in Ephraim & Malah would be to use the continuous output of a modern soft VAD (like Soon et al.'s) to directly control the aggressiveness of the suppression. Implementing this "aggressiveness" control via a simple multiplicative overestimation factor applied to the noise estimate is a known, straightforward design choice for a PHOSITA aiming to achieve stronger attenuation during speech pauses (when speech likelihood is low) and gentler attenuation during speech activity (when speech likelihood is high). This combination directly anticipates the core mechanism described in claim 1 of the '899 patent.

Argument for Obviousness of Claim 15 (Method Claim) and Claim 23 (Computer-Readable Medium)

Claim 15 recites a method that mirrors the system of claim 1. The same arguments for the obviousness of the system in claim 1 apply directly to the steps of the method in claim 15. The combination of Ephraim & Malah, Cohen, and Soon et al. teaches a method of dynamically determining a speech likelihood and using it to compute and apply spectral weighting coefficients where the aggressiveness is adjusted based on speech presence or absence.

Claim 23 claims a non-transitory computer-readable medium containing instructions to perform the method. As it is well-established that implementing a known signal processing method on a computer (via a processor and memory) is obvious, claim 23 would also be rendered obvious by the same combination of prior art that renders claim 15 obvious.

Conclusion

The independent claims of US Patent 10,783,899 appear vulnerable to an obviousness challenge under 35 U.S.C. § 103. The core inventive concept—using a soft speech detector's likelihood output to dynamically control a noise overestimation factor—represents a predictable combination of known elements from the prior art. A person of ordinary skill in the art, faced with the well-known problem of balancing noise reduction with speech preservation, would have been motivated to combine the probabilistic speech detection of Soon et al. with the speech-probability-dependent filtering of Ephraim & Malah, using a straightforward noise overestimation factor as the control mechanism. This combination would have yielded the claimed invention with a reasonable expectation of success.