Document Sample
Apu Powered By Docstoc
					                                      Companding to improve cochlear implants’ speech processing in noise
                                                                                                                          Aparajita Bhattacharya and Fan-Gang Zeng
                                  Hearing and Speech Research Laboratory, Department of Otolaryngology, University of California, Irvine

                                                                                                   (b): Temporal waveform of the consonant /aPa/.                                                                                                   FIGURE 4: Phoneme recognition and Sentence recognition scores for cochlear                                                                                                                  Methods
                                                                                                                                                                                                                                                    implant users.
 The performance of the present day cochlear implants in noise is severely                                                                                                                                                                                                                                                                                                                         Phoneme recognition and sentence recognition tests were conducted in quiet and

  compromised.                                                                                                                                                                                                                                                                                                                                                                                      with steady state speech-shaped noise at varying signal to noise ratios. Stimuli
                                                                                                                                                                                                                                                                                                                                                                 C1                                 were presented at 70 dB SPL and the signal to noise ratios (SNR) varied

 Two tone suppression is thought to be one of the reasons for the outstanding                                                                                                                                                                                                    60
                                                                                                                                                                                                                                                                                                                                                                                                    from -10 dB to 10 dB at 5 dB interval.
                                                                                                                                                                                                                                                                                 40                                                                         0
performance of human ears in noise. Two tone suppression is characterized by a                                                   Without companding                                                     With companding
decrease in response to a tone in the presence of a second tone.                                    Companding changes the temporal waveform.
                                                                                                                                                                                                                                                                                  0                                                                                                                Subjects : Normal : 9 subjects were tested for the vowel recognition tests, 7
                                                                                                                                                                                                                                                                                100                                                                        40
                                                                                                                                                                                                                                                                                      C2                                                                         C2
                                                                                                                                                                                                                                                                                                                                                                                                    subjects for the consonant recognition tests and 10 subjects for the sentence
 Turicchia and Sarpeshkar have shown that companding i.e., compression followed                                                                                                                                                                                                                                                                                                                     recognition tests. Out of these 5 subjects participated in both phoneme
                                                                                                   FIGURE 3: Phoneme recognition and Sentence recognition scores for                                                                                                             40                                                                         0

 by expansion, simulates two tone suppression. Companding enhances the spectral                                                                                                                                                                                                  20                                                                                                                 recognition and sentence recognition tests. Cochlear implant users : 5 Nucleus
                                                                                                   Normal subjects.                                                                                                                                                                                                                                        -20
peaks and therefore reduces the background noise keeping the signal level the                                                                                                                                                                                                                                                                              40                                       24 and 2 Clarion cochlear-implant users also participated in this research.
                                                                                                                                                                                                                                                                                      C3                                                                         C3
same. Their architecture uses a broad filter bank, a compression block, followed                                          100   Vowels                                                                   30    Vowels                                                            60

by a narrowly tuned post filter bank and an expansion block. Both the compression                                          80
                                                                                                                                                                                                         20                                                                      40                                                                         0                                      Tests on normal subjects included two types of stimuli, stimuli processed with the

                                                                                                                                                                                                                                                                                                                               increase in percent score
                                                                                                                           60                                                                                                                                                    20
and expansion blocks consist of an envelope detector and a non-linear unit
                                                                                                                                                                                                                                                                                                                                                                                                    companding strategy alone and stimuli processed with companding strategy
                                                                                                                           40                                                                                                                                                                                                                              40

                                                                                                                                                                                                                                                                percent score
                                                                                                                                                                                                                                                                                100                                                                              C40
followed by a multiplier manipulating the filter output by the non-linear block                                            20                                                                                                                                                           C4
                                                                                                                                                                                                                                                                                           0     5     10    quiet                                                        5         10   quiet
                                                                                                                                                                                                                                                                                                                                                                                                    followed by acoustic simulation of an 8-channel cochlear implant. The phoneme
                                                                                                                                                                                                        -10                                                                      80                                                                        20                 SNR
                                                                                                                                                                                                                                                                                                 X Data
output.                                                                                                                     0
                                                                                                                                                                                                                                                                                 60                                                                                                                 materials included 12 /hvd/ vowels and 20 /aCa/ consonants spoken by a male

                                                                                                                                                                            increase in percent score
                                                                                                                                                                                                              Consonants                                                         40
                                                                                                                                Consonants                                                               30
                                                                                                                                                                                                                -10   -5     0    5    10   Quiet
                                                                                                                                                                                                                                                                                                                                                                                                    talker and a female talker. The target sentence material consisted of 250 HINT
                                                                                                         percent scores

                                                                                                                           80                                                                            20

FIGURE 1: Single channel of the companding architecture.
                                                                                                                                                                                                                                                                                                                                                                                                    sentences spoken by a male talker.
                                                                                                                           60                                                                            10                                                                     100                                                                              C5
                                                                                                                                                                                                                                                                                 80                                                                        20
                                                                                                                           20                                                                           -10
           I/p                                                                                                                                                                                                                                                                                                                                              0
                 F                                                G
                                                                                                                                                                                                        -20                                                                      20                                                                        -20
                                                                                                                          100   Hint sentences                                                           30   Hint sentences
                                                                                                                                                                                                                 -10 -5   0       5    10   Quiet                                 0

                                                                                                                                                                                                                                                                                                                                                           40                                       Companding enhances spectral peaks but also changes the temporal waveforms.
                                                                                                                           80                                                                            20                                                                           C6                                                                         C6
                            ED                C                                     ED   C                                                                                                                                                                                       80                                                                        20
                                                                                                                           60                                                                            10
                     F, G – filters, G is narrowly tuned    ED – envelop detector                                                                                                                                                                                                60
                     C – non-linear compressor unit                                                                        40                                                                             0                                                                      40                                                                                                                Companding does not improve speech performance in normal-hearing subjects
                                                                                                                                                                                                                                                                                 20                                                                        -20
                                                                                                                           20                                                                           -10
                                                                                                                                                                                                                                                                                  0                                                                                                              but it does improve performance in normal subjects listening to an 8 channel
 The purpose of the present study is to investigate the effect of companding                                                0                                                                           -20                                                                                                                                                40
                                                                                                                                                                                                                                                                                       C7 0      5     10    quiet                                               C7 0     5         10   quiet
                                                                                                                                                                                                                                                                                                                                                                                                 cochlear implant simulation, particularly for the vowels.
  architecture on cochlear-implant performance in quiet and with noise.                                                           -10   -5   0     5   10   quiet                                               -10   -5     0    5    10   quiet                                80
                                                                                                                                                                                                                                                                                                 X Data                                                    20                 SNR
                                                                                                                                             SNR                                                                            SNR                                                  60

                                                                                                                                                                                                                                                                                 40                                                                         0

                                                                                                                                without CI simulation, with companding                                                                                                           20
                                                                                                                                                                                                                                                                                                                                                                                                   Companding improves speech performance in cochlear-implant users, particularly
                                                                                                                                                                                                                           without CI simulation
                                                           Results                                                              without CI simulation, without companding
                                                                                                                                with CI simulation, with companding
                                                                                                                                                                                                                           with CI simulation                                     0

                                                                                                                                                                                                                                                                                                                                                                                                    for users with better hearing.
                                                                                                                                with CI simulation, without companding
                                                                                                                                                                                                                                                                                           0     5     10    quiet                                                    0   5         10   quiet
FIGURE 2 (a): FFT of the consonant /aPa/.                                                                                                                                                                                                                                                        SNR                                                                          SNR

                                                                                                                                                                                                                                                                                               vowels with companding
                                                                                                    Maximum benefit from companding is seen for the recognition of vowels.                                                                                                                     vowels without companding
                                                                                                                                                                                                                                                                                                                                                                                                   Future work: We will investigate the performance of companding using a
                                                                                                                                                                                                                                                                                               consonants with companding                                                 consonants
                                                                                                    Improvement in performance is much more prominent with stimuli processed with                                                                                                              consonants without conpanding                                              sentences                 competing talker as a masker. We will analyze the error patterns in consonant
                                                                                                                                                                                                                                                                                               sentences with conpanding
                                                                                                   the companding strategy followed by 8 channel cochlear implant simulation                                                                                                                   sentences without companding                                                                         recognition to find out whether and how companding effects recognition of

                                                                                                   compared to stimuli processed with the companding strategy alone.                                                                                                                                                                                                                                different types of consonants.

                                                                                                    Subjects derived relatively small benefits for recognition of the consonants when
                                                                                                                                                                                                                                                      C1 to C5 are Nucleus 24 users while C6 and C7 are Clarion II users.
                                                                                                   stimuli were processed with companding followed by CI simulation. The scores
 The left panel shows the initial vowel part while the right panel shows that during                                                                                                                                                                  Five of the seven subjects showed significant improvement in scores with                                                                                                        Acknowledgements
                                                                                                   went down for some SNRs when the consonant stimuli were processed with the
  the sudden release of airflow after the stop i.e., the consonant part.                                                                                                                                                                              companding.
                                                                                                                                                                                                                                                                                                                                                                                                 We would like to thank our CI listeners and normal hearing subjects for their time and dedication. This
                                                                                                   companding strategy alone.                                                                                                                                                                                                                                                                    experiment was supported by NIH.
 The green curve shows the data with companding while the red curve shows that                                                                                                                                                                        C6 did not show any improvement with companding. The percent scores for this
                                                                                                    Significant differences between increase in percent scores with and without CI
  without companding.                                                                                                                                                                                                                                 subject are much lower than the rest of the subjects.
                                                                                                   simulation were seen in case of recognition of vowels and sentences.                                                                                                                                                                                                                          • Turicchia L and Sarpeshkar R (2005). “A bio-inspired companding strategy for spectral enhancement.” IEEE
 Significant spectral enhancement seen in case of the initial vowel part. Not much                                                                                                                                                                    Most of the subjects gained little or no improvement in quiet.
                                                                                                    Subjects showed little or no improvement in quiet.                                                                                                                                                                                                                                            transactions on speech and audio processing 2005 Vol 13(2): 243-253
  spectral enhancement seen with companding for the consonant part.

Shared By: