NOISE FLOOR ESTIMATOR
Field of the Invention
The present invention relates to noise estimation in a number of channels in a received audio signal, and more particularly, relates to estimating the noise in any one of the channels in an audio signal received and processed by a cochlea implant system.
Background of the Invention
In an audio processor or speech processor, particularly in relation to hearing implant systems, audio signals are decomposed into sub-bands with each sub-band occupying a different frequency region within the audible spectrum. It is desirable to estimate long term stationary components or noise sources that are contained in the individual channels in real time. Such detection enables the reduction of audible environmental noises, such as car engine noise or fan noise or even various internal body noises presented to a cochlea implant recipient. It also enables the reduction of energy consumed in the electrical stimulation of these noises in the cochlea implant.
The noise sources, particularly in totally implantable cochlea implants, include external noises such as car engines or fan noise, as mentioned previously, or noise sources that are internal to the recipient which may include broadband noises such as heavy breathing, chewing or skin noises as well as narrowband noises such as the recipient's own voice. All these noises have a negative impact on the recipient intelligibility when listening to speech as they can partially or totally mask important parts in the speech spectrum. The degree of the degradation of these perceived signals depends on the instantaneous signal to noise ratio between the speech signal and the noise and the spectral characteristic of the noise itself.
Thus, it is desirable to detect the noise components in the individual channels of the audio processor, which processes the original speech signal, and exclude these from being used in electrical stimulation of the auditory nerves of the implant recipient. At the same time it allows for other channel energies, which contain useful speech information, to be presented for electrical stimulation of the auditory nerves.
Generally the sub-bands are output from a filter bank and may number up to twenty, one for each electrode in the electrode array of the implant. Prior to the signal being presented to the filter bank, it has already been converted from an analogue speech signal into digital form and pre-processed by a filter, such as a DC block filter.
Summary of the Invention
According to a first aspect of the invention there is provided a method of estimating noise floors in an audio signal, the method comprising the steps of: determining a plurality of analysis channels of the audio signal; sampling each one of the analysis channels to determine a respective, sampled magnitude value; comparing each one of the respective, sampled magnitude values with a corresponding one of a current noise floor estimate; counting, for each one of the analysis channels, a number of times the sampled, magnitude value falls outside a predetermined limit for the corresponding one of a current noise floor estimate; and adjusting, for each one of the channels, the corresponding noise floor estimate if the counted number of samples exceeds a corresponding one of a threshold number.
The counting step may comprise counting a number of times a respective sampled, magnitude value is greater than the predetermined limit for the corresponding one of a current noise floor estimate, such that the method further comprises the step of increasing the corresponding one of the current noise floor estimate if the counted number of times the respective sampled, magnitude value is greater than the predetermined limit exceeds the corresponding one of a threshold number. The method may further comprising the step of incrementing a first counter by one each time the respective sampled, magnitude value has a magnitude greater than the predetermined limit for the corresponding one of the current noise floor estimate following the comparing step.
The audio signal may comprise a number of analysis channels, each analysis channel having a bandwidth equivalent to a portion of a speech spectrum, analysed in a cochlea implant system.
The method may further comprise conducting the comparison step over a period of time on a channel-by-channel basis.
The counting step may comprise counting a number of times a respective sampled, magnitude value is less than the predetermined limit for the corresponding one of a current noise floor estimate, such that the method further comprises the step of decreasing the corresponding one of the current noise floor estimate if the counted number of times the respective sampled, magnitude value is less than the predetermined limit exceeds the corresponding one of a threshold number.
The method may further comprise the step of incrementing a second counter by one each time the respective sampled, magnitude value has a magnitude less than the predetermined limit for the corresponding one of the current noise floor estimate after the comparing step. The audio signal may represent a number of analysis channels, each analysis channel having a bandwidth equivalent to a portion of the speech spectrum, analysed in a cochlea implant system.
According to a second aspect of the invention there is provided apparatus for estimating noise floors in an audio signal comprising: means for determining a plurality of analysis channels of the audio signal; means for sampling each one of the analysis channels to determine a respective, sampled magnitude value; comparator means for comparing each one of the respective, sampled magnitude values with a corresponding one of a current noise floor estimate; counter means for counting, for each one of the analysis channels, a number of times the sampled, magnitude value falls outside a predetermined limit for the corresponding one of a current noise floor estimate; wherein for each one of the channels, the corresponding noise floor estimate is adjusted if the counted number of samples exceeds a corresponding one of a threshold number.
The counter means may comprise a first counter for counting a number of times a respective sampled, magnitude value is greater than the predetermined limit for the corresponding one of a current noise floor estimate, such that the corresponding one of the current noise floor estimate is increased if the counted number of times the respective sampled, magnitude value is greater than the predetermined limit exceeds the corresponding one of a threshold number.
The comparator means may be used to determine if the counted number of times the respective sampled, magnitude value is greater than the predetermined limit exceeds the corresponding one of a threshold number. The first counter may be incremented by one each time the respective sampled, magnitude value has a magnitude greater than the predetermined limit for the corresponding one of the current noise floor estimate as determined by the comparator means.
The counter means may comprise a second counter for counting a number of times a respective sampled, magnitude value is less than the predetermined limit for the corresponding one of a current noise floor estimate, such that the corresponding one of
the current noise floor estimate is decreased if the counted number of times the respective sampled, magnitude value is less than the predetermined limit exceeds the corresponding one of a threshold number.
The further comparator means may be used to determine if the counted number of times the respective sampled, magnitude value is less than the predetermined limit exceeds the corresponding one of a threshold number.
According to a third aspect of the invention there is provided a method of estimating the noise floor in an audio signal comprising the steps of: comparing the magnitude of samples of the audio signal with a current noise floor estimate; incrementing a first counter by one for each new sample having a magnitude greater than the current noise floor estimate; incrementing a second counter by one for each new sample having a magnitude less than the current noise floor estimate; further comparing the samples stored in the first counter with a first threshold value and comparing the samples stored in the second counter with a second threshold value.
According to a fourth aspect of the invention there is provided apparatus for estimating the noise floor in an audio signal comprising: first comparator means for receiving the audio signal and comparing a magnitude of samples of the audio signal with a current noise floor estimate; a first counter linked to the first comparator means which is incremented by one for each new sample that has a magnitude greater than the current noise floor estimate as determined by the first comparator means; a second counter linked to the first comparator means for incrementing by one for each new sample that has a magnitude less than the current noise floor estimate as determined by the first comparator means; a second comparator means for comparing the samples stored in the first counter with a first threshold value and for comparing the samples stored in the second counter with a second threshold value.
Brief Description of the Drawings
Preferred embodiments of the invention will hereinafter be described, by way of example only, with reference to the accompanying drawings wherein: Figure 1 is a schematic diagram illustrating the components involved in post processing in an audio processor of a cochlea implant system;
Figure 2 is a schematic diagram showing portions of a noise floor estimator unit; and
Figure 3 is a flow diagram illustrating processes and principle steps in computing a noise floor estimate in a single channel of an audio signal. Figure 4 shows a series of wave forms involved with the signal processing block of an audio processor of a cochlear implant system and a graph of the estimated noise floor.
Detailed Description of a Preferred Embodiment Referring to Figure 1, there is shown a block diagram of part of an audio processor in a cochlea implant. Digital signals that have been converted from the original analogue speech signal via a front end or pre-processor are input to a filter bank 10 and in this particular example, a multi-rate filter bank or MRFB. The filter bank 10 analyses the received digital speech signal and splits the signal into N separate channels. Typically in a cochlea implant system there are 20 or 22 channels. Each channel is designed to have a bandwidth covering part of the spectrum of the received speech signal. Once the signal has been split up into N channels it is input to a further component 20 which makes gain adjustments and detects the envelope of each of the N channels. Each of the N channels are then input to a noise floor estimator 30 in order to estimate the noise floor of the audio channel according to the present invention. Each of the channels are then input to a threshold comparator 40 to enable a comparison to be made between the signal level on each channel and the estimated noise floor and then M out of the N channels are selected by a selection unit 50 on the basis of the comparison. The selected maxima M out of N channels are then forwarded to the loudness growth function (LGF). The selection of maxima on various channels out of the N original channels is the subject of a further invention and a co-pending Australian provisional application No 2003901538 to the same applicant. The primary focus of this invention is to provide an estimation of the noise floor and to adjust that noise floor depending on sampled values of the magnitude of each of the analysis channels. According to the present invention, the audio processor in a cochlea implant system (and more particularly in Totally Implanted Cochlea Implant Systems) employs a noise floor estimation on a channel-by-channel basis which tracks the instantaneous noise floor of each of the analysis channels. The channels are continuously evaluated against these estimates and are considered in the maxima search conducted by the unit
50 only if the magnitude of the channels exceed the current channel noise floor estimate.
Implementation of the noise floor estimator 30 is based on a dual-bin histogram measurement of the monitored magnitudes over time. Each analysis channel measures how often its magnitude lies above and below the current noise floor estimate within a certain time window. This is done by means of two incrementing counters in each analysis channel, as is to be discussed with reference to Figure 2. At the end of the window time, that is Win_T : [samples], the accumulated values of the two incrementing counters are compared against two fixed thresholds that are mutually exclusive to one another. The thresholds are defined as:
Cst_NO_Below_Est:=0.9*Win_T; and
Cst_NO_Above_Est:=0.3 * Win_T.
Referring to Figure 2 there is shown a part of noise floor estimator 30 whereby a channel to be analysed is input to a comparator 60 which compares the magnitude of the signal within that channel with the present noise floor estimate 61. This is performed over a window of time whereby a certain number of samples within the signal are analysed and compared with the current noise floor estimate 61. The comparison is made to find those instances where each samples magnitude values falls outside of a predetermined linit for the current noise floor estimate. In this example, the number of samples having a magnitude greater than the noise floor estimate is flagged and a first counter 62, termed the "Above" counter, is incremented by one for every sample that has its magnitude greater than the noise floor. After the window of time or the time interval has expired, the cumulative values in the counter 62 are compared against the second threshold identified above by a second comparator 64. In a different sampling period or time window, a number of samples of the input channel are compared for their magnitude against the present noise estimate 61 by comparator 60 and for those samples that have their magnitude below the present noise floor estimate, initiates a further counter 63, termed the "Below" counter, to increment by one each time the magnitude of a sample is measured to be below the present noise floor estimate. After the time interval has expired, the cumulative values stored in the counter 63 are compared with the first threshold identified above by comparator 60.
Thus with the case of the incrementing counter 62 that measures the magnitude occurrences that are above the noise estimate, where these exceed the value
Cst_NO_Above_Est a new noise floor estimate is implemented and slowly tracks the increase. More particularly, the noise estimate is increased and may be increased by 16. With regard to the thresholds above, this essentially states that if this is more than 30% of the total number of samples in the integration window, as stored in the counter 62, the new noise floor estimate is increased and is slowly tracked such that:
Noise_Floor_Channel(i) = Noise_Floor_Channel(i) + 2
In the case where the incrementing counter 63, which measures the magnitude of the samples below the present noise estimate, if these exceed the threshold Cst_NO_Below_Est, a new noise floor estimate is implemented and a decrease is rapidly tracked such that:
Noise_Floor_Channel(i) = Noise_Floor_Channel(i) / 2.
Thus in essence, if more than 90% of the total number of samples in the integration window are stored in the counter 63 then a new noise floor which is decreased in value is used as the estimate for the noise floor, and more particularly is made to be half the present value for the noise floor estimate. Thus this is a way of boosting the number of samples having a magnitude greater than the noise floor. This in turn enhances the selection process by the unit 50 to establish the channels having various maxima to be selected to be thereby used in electrical stimulation of the auditory nerves.
With reference to Figure 3 there is shown a flow diagram outlining the various principle steps involved in computing the noise estimate in a single analysis channel. At step 80 the noise floor estimator 30 waits for new channel data to arrive at the comparator unit 60. Once received a determination is made at step 82 as to whether the time interval or window period has elapsed. If not the process moves to step 84 whereby a determination is made as to whether a particular sample has a magnitude greater than the present noise estimate. If yes, the counter 62 is incremented by one at step 86. If no, the counter 63 is incremented by one at step 88. Thereafter the integration time window is decremented by one at step 90 and the process returns to step 80 until the time interval over which the samples are taken has finished. Once this determination is made at step 82 the process moves to step 92 where a determination is
made as to whether counter 63 has more than 90% of all samples sampled within a time interval or window time. If this is the case then the process moves to step 94 where a new noise estimate is implemented which is reduced by a factor of two. The process then moves to step 96 where the counter 63 is reset and the integration window time is also reset. Alternatively, in a separate analysis, if counter 62 has more than 30% of the values sampled in the time interval or window time, then a new noise estimate is increased by 16 and limited to an upper threshold at step 100. The process then moves back to step 96 where counter 62 is reset and the integration window time is reset as well. With reference to Figure 4 there is shown an example of the signals that are involved in the signal-processing block of the audio processor in the cochlear implant. The uppermost waveform 110 shows a raw signal output of an analysis channel assuming that there is a channel response to a signal-frequency signal within the frequency band of this channel. In the waveform 120 there is shown the rectified signal which effectively converts the raw signal in waveform 110 having a bipolar signal amplitude output into a unipolar signal amplitude. The envelope detector 20 smooths out the rectified signal by applying a low pass filter to the signal as shown in the envelope detector output waveform 130. Depicted in graph 140 is the noise floor estimate which includes an increment step size, a decrement factor, an estimator target dynamic range, an estimator attack time and an estimator integration window. The audio sampling interval is the time in which consecutive samples of the audio signal are periodically acquired and is inverse to the audio sampling frequency, which for example amounts to 13,000Hz and is thus around 77 μs.
The estimator attack time is the time within which the noise floor estimator 30 should increase its noise floor estimate for a given estimator target signal dynamic range. This time is generally specified in milliseconds. A typical target or estimator attack time for a long-term stationary noise source is in the order of 200 ms. Within this time the noise floor estimator 30 does increase its estimate from a value practically equal to zero to the level of the anticipated strength of the noise source. The estimator target signal dynamic range is expressed in decibels and specifies the increase in the noise floor estimate for a given estimator attack time. An increment step size is determined by the target dynamic range which optimally matches the anticipated strength of the state-state noise source, divided by the time that the estimator will integrate the noise energy to reach the target or anticipated strength of noise source.
The times that the estimator will integrate the noise source to reach this target equals the target attack time or estimator attack time divided by the integration time window (time period).
Thus the increment step size is given by the following equation:
increment step size = (target signal dynamic range * time period)/ target attack time.
The estimator integration time window involves the noise floor estimator 30 using a gliding window over the rectified signal energies for calculating the noise floor estimate for each analysis channel individually. Within this time it evaluates the incoming signal energy by comparing it with the current noise floor estimate on a channel by channel basis. Two increment counters record the instances of sampled energies which are above or below the current noise floor estimate. At the end of the integration time, a new noise floor estimate is calculated for each analysis channel based on the weighted results of the increment encounters. The new noise floor estimate is increased by the specified increment step size or decreased by the specified decrement step size. The estimator increment step size is the amount by which the current noise floor estimate is increased leading to a new estimate, at the end of the integration time. It is generally expressed as a decimal number. In contrast, the estimator decrement factor is the factor by which the current noise floor estimate is arithmetically divided leading to a new estimate at the end of the integration time. This value is always equal to decimal 2.
The integration time window in turn equals an integer multiple of the channel energy sampling interval, where the integer is chosen in such a way that the integration time window exceeds the quasi-stationary duration of useful signal sources such as speech. Typically, the integration window is between 60 and 100 ms. The channel energy sampling interval is the time in which consecutive samples of the rectified signal energy of the analysis channels are periodically acquired. In one example, this is equal to 32 times the audio sampling interval, and thus amounts to approximately 2.5 ms.
From the above as an example where the integration time window is 60ms, the increment step size can be calculated as follows:
Increment step size = [(24* channel energy sampling interval)*target signal dynamic range]/estimator attack time
An example calculation is where the estimator attack time is chosen to be 250 ms, the channel energy sampling interval is 2.5 ms and the target signal dynamic range equals 128 or equivalently 12 dB. The calculated increment step size is approximately 30.
It will be appreciated by persons skilled in the art that numerous variations and/or modifications may be made to the invention as shown in the specific embodiments without departing from the spirit or scope of the invention as broadly described. The present embodiments are, therefore, to be considered in all respects as illustrative and not restrictive. For example, the method could be readily applied to an audio processor for other hearing prostheses, such as a hearing aid or middle ear mechanical transducer.