Noise Reduction for Hearing Aids: Enabling Communication in Adverse Conditions Rainer Martin October 22, 2013
Communication in Adverse Acoustic Conditions Source: http://cdsweb.cern.ch, accessed on Oct 28, 2012 Introduction Analysis/Synthesis Single Channel NR Multi-Channel Summary Rainer Martin 1 / 48
Hearing Loss Prevalence Data Hearing loss prevalence and hearing aid adoption rates, based on stated hearing loss on the screening survey. Source: http://www.hearingreview.com/issues/articles/2011-02-01.asp Introduction Analysis/Synthesis Single Channel NR Multi-Channel Summary Rainer Martin 2 / 48
Hearing Loss and its Consequences I healthy outer hair cells damaged outer hair cells Source: http://dontlosethemusic.co.nz, accessed on Sept 16, 2013 ◮ Increase of the threshold of hearing • soft sounds are not heard anymore • speech intelligibility (even without additional noise) is insufficient • compensation via strong amplification (up to 70 dB) without exceeding the loudness discomfort level (LDL) • target amplification is derived from fitting rules. Introduction Analysis/Synthesis Single Channel NR Multi-Channel Summary Rainer Martin 3 / 48
Hearing Loss and its Consequences II ◮ Reduction of spectral and/or temporal resolution in the inner ear • speech sounds are loud enough but not intelligible • speech communication in noisy environments is severely degraded • direct compensation of these effects is not possible ◮ Speech enhancement / noise reduction pre-processing is very important for successful rehabilitation! Introduction Analysis/Synthesis Single Channel NR Multi-Channel Summary Rainer Martin 4 / 48
Hearing Aids Sources: Siemens Audiologische Technik, Oticon, varibel Introduction Analysis/Synthesis Single Channel NR Multi-Channel Summary Rainer Martin 5 / 48
Open-Fit Hearing Aids ◮ Open-fit devices • are best for mild to moderate hearing loss with good residual hearing at low frequencies, • are comfortable to wear, • improve own voice reproduction, • require powerful feedback cancellation. Source: www.lloydhearingaid.com Introduction Analysis/Synthesis Single Channel NR Multi-Channel Summary Rainer Martin 6 / 48
Open-Fit Signal Model ◮ Open-fit devices • require very short processing latency, • may be less effective in high levels of ambient noise • a case for active noise control? see [Dalga and Doclo 2013] . noise noise ear speech HA drum speech Introduction Analysis/Synthesis Single Channel NR Multi-Channel Summary Rainer Martin 7 / 48
Wireless Connectivity wireless link ◮ Binaural link for the exchange of settings and parameters ◮ Full audio bandwidth is desired ◮ Audio streaming via wireless relay ◮ Streaming directly from a smartphone to hearing aids ◮ Full bi-directional signal transmission using sensors and computational power of the smartphone Introduction Analysis/Synthesis Single Channel NR Multi-Channel Summary Rainer Martin 8 / 48
Challenges ◮ Users expect effortless communication in complex acoustic environments • many spatially distributed sources • non-stationary, non-Gaussian signals • ambient noise and reverberation • time-varying signal paths • very long impulse responses ◮ This requires optimization of both intelligibility and quality. ◮ Hardware restrictions • very small size of device • very low latency < 10 ms • very low power < 1 mW Sources: Blackberry, Nokia, Siemens, 2010 Introduction Analysis/Synthesis Single Channel NR Multi-Channel Summary Rainer Martin 9 / 48
Outline Introduction 1 Spectral Analysis and Synthesis 2 Single Channel Noise Reduction 3 Multi-channel Speech Enhancement 4 Summary 5
Spectral Analysis and Synthesis Requirements of noise reduction: ◮ High energy compaction of target signal • High spectral resolution of harmonics for voiced speech ⇒ good separation of speech and noise • High temporal resolution for transient sounds ⇒ accurate reproduction of transient speech sounds ◮ High stop-band attenuation ◮ Perfect reconstruction ◮ Low algorithmic delay ◮ High computational efficiency Introduction Analysis/Synthesis Single Channel NR Multi-Channel Summary Rainer Martin 11 / 48
Spectral Analysis / Synthesis ◮ DFT and uniform filter banks, e.g. [Griffin and Lim 1984] • high-resolution • perfect reconstruction • highly efficient ◮ Non-uniform filter banks, e.g. [Hohmann 2002] • resolution according to perceptual model • near-perfect reconstruction ◮ Low-delay filter-bank equalizer, e.g. [Löllmann and Vary 2005], [Vary 2006], [Löllmann and Vary 2008] ◮ Eigenvalue / eigenvector decomposition, e.g. [Ephraim and van Trees 1995] • signal adaptive / optimal • computationally expensive Introduction Analysis/Synthesis Single Channel NR Multi-Channel Summary Rainer Martin 12 / 48
Overlap-Add Analysis and Synthesis segmentation overlap/add y ( k ) � s ( k ) w A w S l l � � � � � Y µ l S µ l noise DFT IDFT reduction To achieve perfect reconstruction the product of these window functions must satisfy the constant-overlap-add constraint ∞ � w A ( n − R ) w S ( n − R ) = 1 k = −∞ where R is the block shift of the windows. Introduction Analysis/Synthesis Single Channel NR Multi-Channel Summary Rainer Martin 13 / 48
Overlap-Add with Symmetric Windows analysis R R x ( n ) n synthesis x ( n ) ˆ n Introduction Analysis/Synthesis Single Channel NR Multi-Channel Summary Rainer Martin 14 / 48
Low Latency Spectral Analysis / Synthesis ◮ Latency is identical to the length of the synthesis window ◮ Use non-symmetric analysis window and short window for synthesis ◮ Family of non-symmetric windows • right-hand side of all analysis and all synthesis windows is identical • left-hand side is variable • use different windows for different speech sounds 0.15 0.15 analysis win synthesis win 0.1 0.1 0.05 0.05 0 0 0 200 400 600 0 200 400 600 DFT: [Mauler and Martin 2007, 2009, 2010], CQT: [Nagathil and Martin, 2012] Introduction Analysis/Synthesis Single Channel NR Multi-Channel Summary Rainer Martin 15 / 48
Low Latency Spectral Analysis / Synthesis with Adaptive Resolution ◮ High spectral resolution required: analysis window synthesis window 1 1 long ana. long syn. 0.5 0.5 0 0 d K−2M−d M M d K−2M−d M M 0 64 128 192 256 320 384 448 512 0 64 128 192 256 320 384 448 512 sample sample ◮ High temporal resolution required: analysis window synthesis window 1 1 short ana. short syn. 0.5 0.5 0 0 d K−2M−d M M d K−2M−d M M 0 64 128 192 256 320 384 448 512 0 64 128 192 256 320 384 448 512 sample sample Introduction Analysis/Synthesis Single Channel NR Multi-Channel Summary Rainer Martin 16 / 48
Adaptive Window Switching analysis R R x ( n ) n synthesis x ( n ) ˆ n Introduction Analysis/Synthesis Single Channel NR Multi -Channel Summary Rainer Martin 17 / 48
Single-channel Noise Reduction segmentation overlap/add y ( k ) s ( k ) � l l � � � � � Y µ l S µ l noise DFT IDFT reduction In the DFT domain we have: � � � � � � ◮ Noisy speech: Y µ l = S µ l + N µ l • frequency index µ • time index l � � � � ◮ Estimated speech coefficient: � S µ l = f ( Y µ l ) Introduction Analysis/Synthesis Single Channel NR Multi -Channel Summary Rainer Martin 18 / 48
Noise Reduction: Basic Tasks � � � � estimation � Y µ l S µ l of speech I D coefficients D F F noise T � � � � T power γ µ l ξ µ l estimation SNR estimation P n ( µ ) : a posteriori SNR : a priori SNR � � � � P n ( µ ) : noise power γ µ l ξ µ l Introduction Analysis/Synthesis Single Channel NR Multi -Channel Summary Rainer Martin 19 / 48
Principle of Single Channel Noise Reduction 20 10 0 power / dB −10 −20 −30 noisy signal spectrum −40 car noise spectrum enhanced spectrum −50 0 1000 2000 3000 4000 frequency / Hz Introduction Analysis/Synthesis Single Channel NR Multi -Channel Summary Rainer Martin 20 / 48
Postprocessing in the Cepstrum Domain for the Reduction of Musical Noise Definition of the real-valued cepstrum : � π � � c y ( q ) = 1 | Y ( e jΩ ) | e jΩq d Ω ln 2 π − π where Y ( e jΩ ) is the spectrum of time domain signal y ( i ) . Some (strange) terminology: cepstrum, quefrency, rahmonic, ... [B.P . Bogert, M.J.R. Healy and J.W. Tukey, 1963] The cepstrum is very well suited to group speech components: ◮ coarse spectral features (envelope), ◮ harmonic structure, and ◮ fine structure of spectrum. Introduction Analysis/Synthesis Single Channel NR Multi -Channel Summary Rainer Martin 21 / 48
Cepstrum of a Voiced Speech Sound signal (f s = 8 kHz) power 300 70 200 60 100 50 dB 0 40 30 −100 20 −200 0 20 40 60 80 100 120 20 40 60 80 100 120 discrete time index DFT bins cepstrum 1.5 1 0.5 0 −0.5 20 40 60 80 100 120 cepstral bins Introduction Analysis/Synthesis Single Channel NR Multi -Channel Summary Rainer Martin 22 / 48
Recommend
More recommend
