{"title": "Improved Silicon Cochlea using Compatible Lateral Bipolar Transistors", "book": "Advances in Neural Information Processing Systems", "page_first": 671, "page_last": 677, "abstract": null, "full_text": "hnproved Silicon Cochlea \n\n\u2022 uSIng \n\nCompatible Lateral Bipolar Transistors \n\nAndre van Schalk, Eric Fragniere,  Eric Vittoz \n\nMANTRA Center for Neuromimetic Systems \n\nSwiss Federal Institute of Technology \n\nCH-IOI5 Lausanne \n\nemail:  vschaik@di.epfl.ch \n\nAbstract \n\nAnalog  electronic  cochlear  models  need  exponentially  scaled  filters. \nCMOS  Compatible  Lateral  Bipolar  Transistors  (CLBTs)  can  create \nexponentially scaled currents when biased using a resistive line with a \nvoltage  difference  between both  ends  of the  line.  Since these  CLBTs \nare  independent  of  the  CMOS  threshold  voltage,  current  sources \nimplemented  with  CLBTs  are  much  better  matched  than  current \nsources  created  with  MOS  transistors  operated  in  weak  inversion. \nMeasurements  from  integrated  test  chips  are  shown  to  verify  the \nimproved matching. \n\n1. INTRODUCTION \n\nSince the original publication of the  \"analog  electronic cochlea\" by  Lyon  and  Mead  in \n1988  [I],  several  other analog  VLSI  models  have  been  proposed  which  try  to  capture \nmore  of  the  details  of  the  biological  cochlear  function  [2],[3],[4].  In  spite  of  the \ndifferences in their design, all these models use filters with exponentially decreasing cut(cid:173)\noff  frequencies.  This  exponential  dependency  is  generally  obtained  using  a  linear \ndecreasing  voltage  on  the  gates  of MOS  transistors  operating  in  weak-inversion.  In \nweak-inversion,  the  drain  current of a  saturated  MOS  transistor depends  exponentially \non  its  gate  voltage.  The  linear  decreasing  voltage  is  easily  created  using  a  resistive \npoly silicon  line;  if there  is  a  voltage  difference  between  the  two  ends  of the  line,  the \nvoltage on the line will decrease linearly all along its length. \n\n\f672 \n\nA. V AN SCHAlK. E.  FRAGNIl1RE. E.  VlrrOZ \n\nThe problem of using MOS  transistors in  weak-inversion as current sources  is  that their \ndrain currents are badly matched. An RMS mismatch of 12% in  the drain current of two \nidentical  transistors  with  equal  gate  and  source  voltages  is  not  exceptional  [5],  even \nwhen  sufficient precautions,  such  as  a  good  layout,  are  taken.  The  main  cause  of this \nmismatch  is  a  variation of the  threshold  voltage  between  the  two  transistors.  Since  the \nthreshold  voltage and  its  variance  are  technology  parameters,  there  is  no  good  way  to \nreduce the mismatch once the chip has been fabricated. \n\nOne can avoid this  problem  using  Compatible Lateral  Bipolar Transistors  (CLBTs)  [6] \nfor  the  current  sources.  They  can  be  readily  made  in  a  CMOS  substrate,  and  their \ncollector current also depends exponentially on their base voltage,  while this  current is \ncompletely  independent  of the  CMOS  technology's  threshold  Voltage.  The  remaining \nmismatch is due to geometry mismatch of the devices, a parameter which is much better \ncontrolled than  the  variance of the  threshold  voltage.  Therefore,  the  use  of CLBTs  can \nyield a  large  improvement in  the  regularity  of the spacing  of the  cochlear  filters.  This \nregularity is  especially important in  a cascade of filters  like the cochlea, since one filter \ncan distort the input signal of all the following filters. \n\nWe have integrated an  analog electronic cochlea as a  cascade of second-order lOW-pass \nfilters,  using CLBTs as exponentially scaled current sources. The design of this  cochlea \nis  based  on  the  silicon  cochlea described  in  [7],  since  a  number  of important  design \nissues, such as stability, dynamic range, device mismatch and compactness, have already \nbeen  addressed  in  this  design.  In  this  paper,  the  design  of [7]  is briefly  presented and \nsome remaining possible improvements are identified. These improvements, notably  the \nuse  of Compatible Lateral  Bipolar Transistors as current sources,  a  differentiation  that \ndoes  not  need  gain  correction  and  temperature  independent  biasing  of  the  cut-off \nfrequency, are then discussed in more detail. Finally,  measurement results of a  test chip \nwill be presented and compared to the design without CLBTs. \n\n2. THE ANALOG ELECTRONIC COCHLEA \n\nThe basic building  block for  the  filters  in  all  analog  electronic  cochlear  models  is  the \ntransconductance amplifier, operated in  weak inversion. For input voltages smaller than \nabout 60 mV pp,  the amplifier can be approximated as a linear transconductance: \n\n(1) \n\nwith transconductance gm given by: \n\n10 \n\ngm = 2nUT \n\n(2) \nwhere Io  is  the  bias current,  n  is  the  slope factor,  and the  thermal  voltage  UT = kT/q = \n25.6 mV at room temperature. \n\nThis  linear range  is  usually  the  input range  used  in  the  cochlear filters,  yielding  linear \nfilters.  In  [7],  a  transconductance  amplifier  having  a  wider  linear  input  range  is \nproposed.  This  allows  larger  input  signals  to  be  used,  up  to  about  140 m Vpp. \nFurthermore,  the  wide  range  transconductance  amplifier  can  be  used  to  eliminate  the \nlarge-signal instability shown to  be present in  the original second-order section [7]. This \nsecond-order section will be discussed in  more detail in section 3.2. \n\n\fImproved  Silicon  Cochlea  Using  Compatible  Lateral  Bipolar Transistors \n\n673 \n\nThe traditional  techniques  to  improve matching  [5],  as  for  instance larger device  sizes \nfor  critical  devices  and  placing  identical  devices  close  together  with \nidentical \norientation, are also discussed in  [7]  with respect to  the implementation of the  cochlear \nfilter  cascade.  The  transistors  generating  the  bias  current  10  of the  transconductance \namplifiers in  the second-order sections were identified as the most critical devices,  since \nthey  have  the  largest  effect  on  the  cut-off  frequency  and  the  quality  factor  of  each \nsection.  Therefore,  extra  area  had  to  be  devoted  to  these  bias  transistors.  A  further \nimprovement  is  obtained  in  [7]  by  using  a  single  resistive  line  to  bias  both  the \ntransconductance amplifiers controlling the cut-off frequency  and the  transconductance \namplifier controlling the quality factor. The quality  factor Q is then changed by varying \nthe source of the  transistor which  biases  the  Q  control amplifier.  Instead  of using  two \ntilted resistive  lines,  this  scheme uses  only  one tilted resistive  line  and a  non-tilted  Q \ncontrol line, and therefore doesn't need to rely on an identical tilt on both resistive lines. \n\n3. IMPROVED ANALOG ELECTRONIC COCHLEA \n\nThe design discussed in  the previous section already showed a substantial improvement \nover  the  first  analog  electronic  cochlea  by  Lyon  and  Mead.  However,  several \nimprovements remain possible. \n\n3.1 VT VARIATION \n\nThe  bias  transistors  have  been  identified  as  the  major  source  of  mismatch  of  the \ncochlea's parameters. This mismatch is mainly due to variation of the  threshold voltage \nVT  of the  MOS  transistors.  Since  the  drain  current  of a  saturated  MOS  transistor  in \nweak-inversion depends exponentially on  the difference between its gate-source voltage \nand its  threshold  voltage,  small variations in VT  introduce large variations in  the  drain \ncurrent of these transistors, and since both the cut-off frequency and the quality factor of \nthe  filters  are  proportional  to  these  drain  currents,  large  parameter  variations  are \ngenerated  by  small  V T  variations.  This  problem  can  be  circumvented  by  the  use  of \nCMOS Compatible Lateral Bipolar transistors as bias transistors. \n\nA  CMOS  Compatible  Lateral  Bipolar  Transistor  is  obtained  if  the  drain  or  source \njunction of a  MOS  transistor is forward-biased  in  order  to  inject minority  carriers  into \nthe local substrate. If the gate voltage is negative enough (for an n-channel device),  then \nno current can flow  at the surface and the operation is purely bipolar  [6].  Fig.  1 shows \nthe major flows of current carriers in  this mode of operation, with  the source,  drain and \nwell terminals renamed emitter E, collector C and base B. \n\nISub \n\n-... \nholes \n\nVBC<O \n\n:fG \n\nC \n\n........  electrons \n\np \n\nn \n\n~ \n\nFig.  1.  : Bipolar operation of the MOS  transistor: carrier flows and symbol. \n\n\f674 \n\nA.  V AN  SCHAlK. E.  FRAGNIERE. E. VITIOZ \n\nSince there  is no  p+ buried layer  to  prevent  injection  to  the  substrate,  this  lateral  npn \nbipolar  transistor  is  combined  with  a  vertical  npn.  The  emitter current  IE  is  thus  split \ninto a base current IB,  a  lateral collector current Ic and a substrate collector current Isub\u2022 \nTherefore,  the  common-base current gain ex.  = -IdlE cannot be close  to  1.  However,  due \nto the very small rate of recombination inside the well and to  the high emitter efficiency, \nthe common-emitter current gain ~ = IeIlB can be large. Maximum values of ex.  and ~ are \nobtained  in  concentric  structures  using  a  minimum  size  emitter  surrounded  by  the \ncollector and a minimum lateral base width. \n\nFor VCE = VBE-VBC larger than a few  hundred millivolts,  this transistor is in active mode \nand the collector current is given, as for a normal bipolar transistor, by \n\nfu \nk=~e~ \n\nW \n\nwhere ISb  is the specific current in bipolar mode, proportional to  the cross-section of the \nemitter  to  collector  flow  of carriers.  Since  k  is  independent  of  the  MOS  transistor \nthreshold voltage V T,  the main source of mismatch of distributed MOS current sources is \nsuppressed, when o....BTs are used to create the current sources. \n\nVC.c  D \n\n__ ...... --'=-_B \nc::::J \n\n0+ \n\n_ \np+ \n\nlEI \npoIy-Si \n(b) \n\nFig. 2. o....BT cascode circuit (a) and its layout (b). \n\nA  disadvantage of the  CLBT is  its  low  Early  voltage,  i.e.,  the  device has  a  low  output \nresistance.  Therefore,  it is  preferable  to  use  a  cascode  circuit as  shown  in  fig.  2.  This \nyields  an  output resistance several hundred  times  larger than  that of the  single  o....BT, \nwhereas the area penalty, in a layout as shown in fig 2b, is acceptable. \n\nAnother disadvantage  of the  CLBTs,  when  biased  using  a  resistive  line,  is  their  base \ncurrent,  which  introduces  an  additional  voltage  drop  on  the  resistive  line.  However, \nsince the  cut-off frequencies  in  the  cochlea  are  controlled  by  the  output current of the \nCLBTs  and  since  these  cut-off frequencies  are  relatively  small,  typically  20  kHz,  the \noutput current  of the  CLBTs  will  be  small.  If the  common-emitter  current  gain  ~ is \nmuch larger than 1,  the base current of these o....BTs will be very small,  and the voltage \nerror  introduced by  the  small  base  currents  will  be  negligible.  Furthermore,  since  the \ncut-off frequencies  of the  cochlea  will  typically  span  2  decades  with  an  exponentially \ndecreasing cut-off frequency from the beginning to  the end, only the first few filters  will \nhave any noticeable influence on the current drawn from the resistive line. \n\n3.2 DIFFERENTIATION \n\nThe  stabilized  second-order  section  of  [7]  uses  two  wide  range  transconductance \namplifiers  (A 1 and A2 in  fig.  3)  with  equal  bias  current and equal  capacitive  load,  to \ncontrol  the  cut-off  frequency.  A  basic  transconductance  amplifier  (A3)  is  used  in  a \n\n\fImproved Silicon  Cochlea  Using  Compatible  Lateral  Bipolar Transistors \n\n675 \n\nfeedback path to  control the quality factor of the  filter. The voltage VOU1  at the output of \neach second-order stage represents the basilar membrane displacement. Since the output \nof the  biological  cochlea  is  proportional  to  the  velocity  of the  basilar  membrane,  the \noutput of each second-order stage has to be differentiated. In  [7]  this is done by creating \na copy  of the  output current Lru- of amplifier A2  at every  stage.  Since  the  voltage  on  a \ncapacitor  is  proportional  to  the  integral  of  the  current  onto  the  capacitor,  Idit  is \neffectively proportional to  the  basilar membrane velocity.  Yet,  with equal  displacement \namplitudes,  velocity  will be much larger for  high  frequencies  than  for  low  frequencies, \nyielding  output  signals  with  an  amplitude  that  decreases  from  the  beginning  of  the \ncochlea to  the end. This can be corrected by normalizing Lru- to  give equal amplitude  at \nevery output. A second resistive line with identical tilt controlling the gain of the current \nmirrors  that  create  the  copies  of  Idit  at  each  stage  is  used  for  this  purpose  in  [7]. \nHowever, if using a single resistive line for the control of the cut-off frequencies  and the \nquality factor  improves the performance of the  chip,  the  same is  true  for  the  control of \nthe current mirror gain. \n\nfromprev. \nsection \n\nFig. 3. One section of the cochlear cascade, with differentiator. \n\nAn  alternative  solution,  which  does  not  need  normalization,  is  to  take  the  difference \nbetween VOuI  and VI  (see fig.  3).  This can be shown  to  be equivalent to  differentiating \nV Out. with OdB  gain at the cut-off frequency for all stages. This can be easily done with a \ncombination  of 2  transconductance  amplifiers.  These  amplifiers  can  have  a  large  bias \ncurrent,  so they  can  also be used  to buffer the  cascade voltages before connecting  them \nto  the  output pins  of the  chip,  to  avoid  charging  the  cochlear  cascade  with  the  extra \ncapacitance introduced by the output pins. \n\n3.3 TEMPERATURE SENSITIVITY \nThe cut-off frequency of the first and the last low-pass filter in the cascade can be  set by \napplying voltages to  both ends of the resistive line, and the intermediate filters  will have \na cut-off frequency  decreasing  exponentially  from  the  beginning  to  the  end. Yet,  if we \napply  directly  a  voltage  to  the  ends  of the  resistive  line,  the  actual  cut-off  frequency \nobtained will depend on the temperature, since the current depends exponentially on the \napplied  voltage  normalized  to  the  thermal  voltage Ur  (see(3).  It  is  therefore  better  to \ncreate the  voltages  at both  ends  of the  resistive  line  on-chip  using  a  current  biasing  a \nCLBT with its base connected to its collector (or its drain connected to its gate if aMOS \ntransistor  is  used).  If this  gate  voltage  is  buffered,  so  that  the  current  through  the \nresistive line  is  not drawn from  the  input current,  the  bias currents of the  first  and  last \nfilter, and thus the cut-off frequency of all filters can be set, independent of  temperature. \n\n\f676 \n\nA.  V AN SCHAlK, E.  FRAGNIERE, E.  VITTOZ \n\n3.4 THE IMPROVED SILICON COCHLEA \n\nThe improved  silicon  cochlea is shown in  figure 4.  It uses  the cochlear sections shown \nin figure 3, CLBTs as  the bias transistors of each filter,  and one resistive line  to  bias all \nCLBTs. The resistive line is biased  using  two bipolar current mirror structures and two \nvoltage buffers, which allow  temperature  independent biasing  of the  cut-off frequencies \nof the  cochlea.  A  similar  structure  is  used  to  create  the  voltage  source  V q  to  control, \nindependent of temperature, the actual quality factor of each section. The actual bipolar \ncurrent mirror implemented uses the cascode structure shown in  figure 2a,  however this \nis not shown in figure 4 for clarity. \n\nVdiffl \n\nFig 4. The improved silicon cochlea \n\n4. TEST RESULTS \n\nThe proposed silicon cochlea has been integrated using the ECPD15  technology at ES2 \n(Grenoble,  France),  containing  104  second-order  stages,  on a  4.77mm  X  3.21mm die. \nEvery second stage is connected to a pin, so its output voltage can be measured. In fig.  5, \nthe  frequency  response curves after on-chip derivation  are shown  for  the  output taps  of \nboth  the cochlea described in  [7]  (left),  and our  version  (right). This  clearly  shows  the \nimproved regularity of the cut-off frequencies  and  the  gain obtained using CLBTs. The \ndrop-off in  gain  for  the  higher frequency  stages  (right)  is  a  border effect,  since  at  the \nbeginning of the  cochlea no accumulation  of gain has  yet taken  place. In  the  figure  on \nthe left this is not visible, since the first nine outputs are not presented. \n\n-20 \n\n\u00b730 \n\n10 \n\n~ \n~ 0 \n\u00b710 \n\n\u00b720 \n\n\u00b730 \n\nF~(Hz) \n\nF~(Hz) \nFig.5. Measured frequency responses at the different taps. \n\n10000 \n\n20000 \n\nIn  fig . 6  we  show  the  cut-off frequency  versus  tap  number  of both  chips.  Ideally,  this \nshould  be  a  straight  line  on  a  log-linear  scale,  since  the  cut-off  frequency  decreases \n\n\fImproved  Silicon  Cochlea  Using  Compatible  Lateral  Bipolar Transistors \n\n677 \n\nexponentially  with  tap  number.  This  also  clearly  shows  the  improved  regularity  using \nCLBTs as current sources. \n\nlOOOO\u00b7r-------------------, \n\n10 \n\n15 \n\n20 \n\n25 \n\n30 \n\n200\u00b7~----------------~~ \n\no \n\n10 \n\n20 \n\n30 \n\n40 \n\n50 \n\nFig.6. Cut-off frequency (Hz) versus tap number for both silicon cochleae. \n\n5. CONCLUSIONS \n\nSince the biological cochlea functions as a distributed filter,  where the natural frequency \ndecreases exponentially with the position along the basilar membrane, analog electronic \ncochlear models  need  exponentially  scaled  filters.  The  output current of a  Compatible \nLateral  Bipolar  Transistor  depends  exponentially  on  the  base-emitter  voltage.  It  is \ntherefore easy to  create exponentially scaled current sources using CLBTs biased with a \nresistive polysilicon line. Because the CLBTs are insensitive to  variations of the CMOS \nthreshold  voltage  VT,  current  sources  implemented  with  CLBTs  are  much  better \nmatched than current sources using MaS transistors in weak inversion. \n\nRegularity  is  further  improved  using  an  on-chip  differentiation  that  does  not  need  a \nsecond resistive line to correct its gain, and therefore doesn't depend on identical tilt on \nboth  resistive  lines.  Better independence of temperature  can  be  obtained  by  fixing  the \nfrequency domain of the cochlea using bias currents instead of voltages. \n\nAcknowledgments \n\nThe  authors  would  like  to  thank  Felix  Lustenberger  for  simulation  and  layout  of the \nchip. We are also indebted to Lloyd Watts for allowing us to  use his measurement data. \n\nReferences \n\n[1]  R.F.  Lyon  and  C.A.  Mead,  \"An  analog  electronic cochlea,\" IEEE  Trans.  Acoust .\u2022 \n\nSpeech.  Signal Processing, vol. 36, pp. 1119-1134, July  1988. \n\n[2]  R.F. Lyon,  \"Analog implementations of auditory models,\" Proc.  DARPA  Workshop \n\nSpeech and Natural Language. San Mateo, CA:Morgan Kaufmann, 1991. \n\n[3]  W.  Liu,  et.  al.,  \"Analog  VLSI  implementation  of an  auditory  periphery  model,\" \n\nAdvances Res.  VLSI,  Proc.  1991 Santa Cruz Con/., MIT Press, 1991, pp. 153-163. \n\n[4]  L.  Watts,  \"Cochlear  Mechanics:  Analysis  and  Analog  VLSI,\"  Ph.D.  thesis, \n\nCalifornia  Institute of Technology, Pasadena, 1992. \n\n[5]  E.  Vittoz,  \"The  design  of  high-performance  analog  circuits  on  digital  CMOS \n\nchips,\" IEEE 1.  Solid-State Circuits, vol. SC-20, pp. 657-665, June 1985. \n\n[6]  E.  Vittoz,  \"MaS  transistors  operated  in  the  lateral  bipolar  mode  and  their \napplication  in  CMOS  technology,\"  IEEE 1.  Solid-State  Circuits,  vol.  SC-24,  pp. \n273-279, June 1983. \n\n[7]  L. Watts, et. al.,  \"Improved implementation of the silicon cochlea,\" IEEE 1.  Solid(cid:173)\n\nState Circuits, vol.  SC-27, pp. 692-700, May 1992. \n\n\f", "award": [], "sourceid": 1173, "authors": [{"given_name": "Andr\u00e9", "family_name": "van Schaik", "institution": null}, {"given_name": "Eric", "family_name": "Fragni\u00e8re", "institution": null}, {"given_name": "Eric", "family_name": "Vittoz", "institution": null}]}