As for the Noise Tracker II, quantity of gain reduction of low frequency and the high frequency is decided by the possibility that there is a speech. Which frequency is this speech probability decided with reference to?
Is it specific frequency?
Is it average of the specific frequency?
Is it average of all frequency?
By Jenny Groth
It is perhaps easier if I just list the steps that happen in the NoiseTracker II processing. For background, the quantity of gain reduction when speech is present is based on the speech importance function published by Pavlovic (1994, Ear & Hearing).
- Analyze the incoming sample of sound for the presence of speech
- If no speech is present, update the estimate of the noise in the 17 warped frequency bands
- Subtract noise estimate in each band from total signal in each band. This gives the estimated signal-to-noise ratio in each band
- Look at probability of speech (this is information that we got in step 1). If probability is zero (no speech) apply gain reduction in each frequency band based on estimated signal-to-noise ratio that we got in step 3
- If probability of speech is higher than zero, apply speech importance weighting to gain reduction.
Example: let’s say that NoiseTracker II is set to “moderate”. Then you can have a maximum gain reduction of 6 dB at all frequencies if there is a very poor signal-to-noise ratio at each of these frequencies (as shown by the green curve in the graph below). Let’s look at the band around 1kHz. If probability of speech is zero and signal-to-noise ratio is determined to be 0 dB, then gain reduction might be around 4 dB according to the gain reduction function (draw a vertical line up from 0 dB SNR to the green curve and you will see that it is about -4 dB on the vertical axis). If, however, speech probability is found to be 30%, and signal-to-noise ratio in this band is 0 dB, then the speech weighting is applied. From the red curve in your email below, we can see that gain will be added back on at 1 kHz when speech probability is 30%. There are no numbers on the y-axis, but let’s say it is 2 dB. This means that the final gain reduction would be:
-4dB + 2 dB = -2 dB. The -4dB is the reduction that was based on the signal-to-noise ratio, while the +2dB is what gets added back on in this band because the speech probability is 30%. If we took this same example for a band around 250 Hz, the gain reduction would end up being more than -4 dB.
We calculate speech probability on an ongoing basis and use it for different things in the hearing aid processing, like the Environmental Classifier and the Environmental Optimizer, as well as NoiseTracker II. The data used by the speech probability estimator includes modulation and overall level of the sound, as well as the presence of temporal patterns that would be typical of speech (alternating high and low frequency energy). This data is fed to a model that delivers a probability that the signal contains speech. So it is not only looking at a particular frequency or an average of anything, but rather a characteristic pattern in time of the energy distribution across frequencies.