Sunitha.S.L is working as Professor in BGS Institute of Technology, B G Nagara-571448.The research paper published by IJSER journal is about Adaptive Complex Transformation for Sensorineural Impairment: A Practical Approach 1
ISSN 2229-5518
Adaptive Complex Transformation for
Sensorineural Impairment: A Practical Approach
Sunitha.S.L and V.Udayashankara
Abstract— Hearing impairment is the number one chronic disability affecting many people in the world. Background noise is particularly damaging to speech intelligibility for people with hearing loss especially for sensorineural loss patients. Several investig ations on speech intelligibility have demonstrated that sensorineural loss patients need 5-15 dB higher SNR than the normal hearing subjects. This paper describes a practical approach using adaptive complex transformation filtering for sensorineural impairment to improve the SN R of the speech signal. The computer simulated results show superior convergence characteristics of the adaptive complex transformation algorithm by improving the SNR at least 7dB for input SNR‘s less than and equal to 0 dB, with 120 convergence ratio, better t ime and frequency characteristics.
Index Terms— Hearing Impairment, Adaptive filter, Sensorineural loss, complex transformation and SNR improvement.
—————————— ——————————
EARING impairment is the preamble chronic disability, affecting people in the world. Many people have great difficulty in understanding speech with background noise. This is especially true for a large number of elderly peoples and
sensorineural impaired persons.
Hearing loss or deafness can be broadly classified in- to 2 types. Conductive loss: This type of hearing disability can be measured by audiograms and the intelligibility of the signal can be easily resorted by amplification.
Sensorineural loss: This is a broad class of hearing impairments its origin is in the cochlea or auditory nervous system. sensorineural loss disorders are difficulty to remedy. This type of defects may be due to congenital or hereditary factors, disease, tumors, old age, long-term exposure to indus- trial noise, acoustic trauma or the action of toxic agents etc. The sensorineural loss patient‘s experiences difficulty in mak- ing fine distinction between speech sounds, particularly those having a predominance of high frequency Energy [6], [10]. He may hear the speaker‘s voice easily, but be unable to distin- guish. For example between the words ‗fat‘ and ‗sat‘ [7], [15]. Two features of sensorineural impairment particularly detri- mental to the perception of speech are high tone loss and compression of the dynamic range of the ear. A high tone loss is analogous to low pass filtering. Amplification of the high tones may improve intelligibility, but in these circumstances dynamic range of the ear is a handicap [9], [4]. Because, the dynamic range of the impaired ear may not be sufficient to accommodate the range of intensities in speech signals. So, the stronger components of speech are perceived at a level, which is uncomfortably loud, while the weaker components are
————————————————
Sunitha.S.L is working as Professor in BGS Institute of Technology, B G Nagara-571448.
E-mail: sunithanov27@yahoo.co.in
V.Udayashankara is working as Research Scientist in JSS Re- search Foundation, Mysore.
E-mail: v_udayashankara@yahoo.co.in
not heard at all [2], [4]. Most of the defects in transmission chain up to cochlea can be successfully rehabilitated by means of surgery. The great majority of the remaining inoper- able cases are sensorineural hearing impaired patients [3]. Digital technology has made an important contribution in the field of audio logy. Digital signal processing methods offer great potential for designing a hearing aid but, today‘s Digital Hearing Aid are not up to the expectation for sensorineural loss patients. Hearing-impaired patients applying for hearing aid reveal that more than 50% are due to sensorineural loss. So for only direct Adaptive filtering methods are suggested in the literature for the minimization of noise from the speech signal for sensorineural loss patients [7], [13].
Adaptive NLMS noise canceller provides SNR im- provement, with less complexity and is having the capability to track the non-stationary environment. But they are having poor convergence performance. Hence they need more time to converge into the optimal solution and become less feasible in real time applications for digital hearing aid [1], [11] & [14].
Convergence speed of time domain LMS adaptive fil- ters depends on the ratio of the maximum to minimum eigen- values of the input autocorrelation matrix. Filters having inputs with wide eigenvalue spread requires longer time to converge. Convergence performance of the standard LMS algorithm can be improved by using frequency domain filtering [16]. This type of adaptive filter is called as frequency domain adaptive filter or transform domain adaptive LMS (TDLMS) filter [5], [8]. In this paper, TLMS is implemented by using Running DFT- LMS to reduce the computational complexity of FFT-LMS.
The transformation in the transform domain LMS filter can be implemented in a variety of ways. This transform is
―continuous flow‖ transformation and therefore computations
can be reduced. The LMS spectrum analyzer is an adaptive
IJSER © 2012
The research paper published by IJSER journal is about Adaptive Complex Transformation for Sensorineural Impairment: A Practical Approach 2
ISSN 2229-5518
system that can be used for the calculation of DFT-LMS as shown in Fig.1.
Fig. 1. LMS spectrum analyzer-LMS
EIGENVALUE RATIO AND TIME PLOTS
The algorithm is evaluated for corrupted speech sig- nals with different types of noises like cafeteria, low frequency and babble noise with different SNR. The input signal is rec- orded with sampling frequency 22050 Hz in different noisy conditions to evaluate the performance of the algorithm. For different input SNR, the output SNR and eigenvalue ratios are calculated as shown in Table 1. Results show significant im- provement in convergence performance by reducing the ei- genvalue ratio to 120.09 and 6.7 dB output SNR improve- ments for 0dB input SNR. Fig. 2 shows the time plots for pure signal, corrupted signal (input signal) with –5dB SNR, and the FFT-LMS filtered signal. Fig. 3 shows the autocorrelation of the corrupted signal after DFT.
The eigenvalue ratio of complex-LMS is less com-
pared to NLMS methods. Hence the convergence perfor- mance of the algorithm is significantly improved. Table 1 show
that the noise is reduced from the corrupted signal and the
At time lN , W
k
WlN
becomes.
lN 1
X m am
m lN N
lN 1
speech quality is also improved. The eigenvalue distribution of the input auto correlation matrix has been derived after DFT and power normalization.
X m m lN N
lN 1
1
m lN N
N
lN 1
X m e
j 2 m / N
m lN N
X m e
j 2 ( N 1) m / N
1
W is proportional to the DFT of the input signal X
k
1
. Thus, if we set
, the weight vector W
2 k
will be exactly
Fig. 2. Original, contaminated and filtered signal for adaptive
proportional to the DFT of the previous N samples of the in- put X k at times k that are integer multiples of N as shown in Fig 1. The remaining part will performer as an adaptive filter. The function of LMS spectrum analyzer-LMS is same as con- ventional FFT-LMS with less computational complexity.
The performance of the algorithm has been evaluated using output SNR, eigenvalue ratio, time plots and intelligibility tests.
complex transformation algorithm LMS spectrum analyzer-LMS
TABLE 1
SNR of input and output signals.
IJSER © 2012
The research paper published by IJSER journal is about Adaptive Complex Transformation for Sensorineural Impairment: A Practical Approach 3
ISSN 2229-5518
TABLE 4
Intelligibility improvements by complex transformed
LMS for three groups of subjects.
Fig. 3. Autocorrelation of the input corrupted signal after DFT.
In order to measure the performance of clinical intelli- gibility tests of the algorithms, listening tests were carried out. The tests were conducted on both hearing impaired and nor- mal hearing persons. The experiment was carried out in a room whose size was about 4 m by 5 m. The room was car- peted but no attempt was made to improve the room acoustics otherwise. The main speaker and the noise source were placed 2.5 feet away from the microphones. For speech intel- ligibility test, we processed 10 sentences with different noise. These tests were performed on 15 subjects, 5 with normal hearing (Group 1), 5 with a mild to moderate SNHL (Group 2) and 5 with moderate to severe SNHL loss (Group 3).
In the experimental evaluation, the target source was
a male speaker reading sentences and interference consisted of 3 different types of noise (1) cocktail party noise (2) five speaker babble (3 male and 2 female) (3) low frequency noise. The noise level is varied to get different SNR. The subjects were listened the original, the noisy and the filtered signals. The percentage of correct responses was recorded. The re- sults are displayed in Tables 2, 3 and 4 for –5dB input SNR. The results indicate that a considerable improvement is ob- tained, particularly for moderate to severe SNHL subjects. Filter shows reduced average intelligibility of 4 % with normal subjects, 13 % with mild to moderate SNHL subjects and 9 % with moderate to severe SNHL subjects as compared to NLMS with cocktail party noise.
TABLE 2
Average intelligibility score for the noiseless signal
Group1 | Group 2 | Group 3 |
96 % | 78 % | 63 % |
TABLE 3
Average intelligibility score for the signal plus noise
Running DFT-LMS method can be used for noise re- duction in speech signals. This algorithm is excellent com- pared to NLMS and single source NLMS algorithm in terms of convergence performance. The eigenvalue ratio is 120 for zero dB and is very less compared to time domain adaptive methods. Hence, this complex transformed adaptive filter can quickly converge to the optimal solution. Off line tests under different conditions show output SNR improvement is lesser than NLMS methods.
[1] Bernard W idrow and Samuel D.Stearns. Adaptive Signal Processing, Pearson Education Asia, 2002.
[2] Baer.T., Moore.B.C., Kulk.K., Effects of low pass filtering on the intel- ligibility of speech in noise for people with and without dead regions at high frequencies. Journal on Acoustic soc Am. 2009 Sept: 112(3 pt 1). page No 1133-44.
[3] Baer. T., Moore.B.C and Gatechouse. Spectral contrast enhancement of speech in noise for listeners with sensorineural hearing impairment: effects on intelligibility, quality, and response times. Journal on Rehabilita- tion, 1993:30(1). Page No 49-72.
[4] Chung.K. Challenges and recent developments in Hearing Aids, Trends Amplif.2009:8(3). Page No 83-124.
[5] Francosie Beaufays, Transform domain adaptive filters: An analytical approach, IEEE Trans. on Signal processing, Vol 43 no2. Feb 1995.
[6] Hornsby.B.W and Ricketts.T.A. The effects of compression ratio, signal-to-noise ratio and level on speech recognition in normal-hearing listeners. Journal on Acoustics soc Am. 2001 June: 109(6). Page No
2964-73.
[7] Dr. Harry Levitt, Noise reduction in Hearing aids: An overview. Journal of Rehabilitation Research and Development. Jan-Feb 2001,Vol.38. No.1. [8] Mohammed A Shamma. Improving the speed and performance of adaptive equalizers via transform based adaptive filtering. IEEE Transac- tions on Signal processing. 2591 Ashurst Rd. University Heights, Ohio,
44118, USA.
[9] Moore. B.C., Stainsby . T.H., Alcantara. J.I., Kuhnel. V, The effect of speech intelligibility of varying compression time constants in a digital hearing aid. International Journal on Audio logy. 2004 Jul-Aug: 43(7), Page No. 399-409.
[10] Rankovic.C.M, Factors governing speech reception benefits of adap- tive linear filtering for listeners with sensorineural loss. Journal on Acous- tics Soc Am. 1998 Feb: 103(2). Page No 1043-57.
[11] Shaul Florian and Neil J Bershad. A Weighted Normalized Frequency
Domain LMS Adaptive Algorithm. IEEE Transactions on Acoustics speech
IJSER © 2012
The research paper published by IJSER journal is about Adaptive Complex Transformation for Sensorineural Impairment: A Practical Approach 4
ISSN 2229-5518
and signal processing, July 1998, Vol36, no 7.
[12] Shanks. J.E., W ilson. R.H.Larson, W illiams. D., Speech recognition performance of patients with sensorineural hearing loss under unaided and aided conditions using linear and compression hearing aids. Ear hear. 2002 Aug:23(4), page No. 280-90.
[13] Shields.P.W and Campbell.B.R. Improvements in intelligibility of noisy reverberant speech using a binaural sub band adaptive noise-cancellation processing scheme. Journal on Acoustics Soc Am. 2001Dec: 110(6). Page No 3232-42.
[14] Simon Haykin, Adaptive Filter Theory, Pearson Education Asia, 4th
Edition, 2002.
[15] V.Udayashankara and A.P.Shivaprasad., Digital Hearing Aid: A Review. World congress on Medical physics & Biomedical Engineering. Brejil, Aug. 1994. Page No 21-26.
[16] V.Udayashankara , A.P.Shivaprasad. The application of voltera LMS Adaptive filtering to speech enhancement for the Hearing Impairment.
4th Euro-speech conference on speech communication and Technology,
Sept.1995, Mandrid, Spain.. Page No 1 No 91-8.
IJSER © 2012