ISCA Archive
`http://www.iscaĆspeech.org/archive
`
`First European Conference on
`Speech Communication and Technology
`EUROSPEECH '89
`Paris, France, September 27Ć29, 1989
`
`Multi-DSP and VQ-ASIC Based Acoustic Front-End
`for Real-Time Speech Processing Tasks
`
`Abdulmesih Aktas and Harald Höge
`
`Siemens AG, München
`Otto-Hahn-Ring 6, D-8000 München 83, West Germany
`
`Abstract
`
`This paper describes the architecture of a multi(cid:173)
`DSP based acoustic front-end (AkuFE) developed
`within the speaker adaptive continuous speech
`understanding and dialogue system SPICOS1. The
`AkuFE
`is designed as a configurable high
`performance signal processing VMEbus system
`employing up
`to
`five Texas
`Instruments
`TMS320C25 signal processors and an ASIC for
`vector quantization (VQ) developed within the
`project. The system is realized on three boards and
`achieves a total computational power of more than
`100 MIPs. In this case two VQ processors can work
`in parallel. A 68020 based workstation serves as
`host computer. The AkuFE is employed for the
`real-time acoustic-phonetic decoding task in the
`SPICOS system. Due to its flexibility, it can be
`used for a wide range of real-time speech
`processing tasks.
`
`1. Introduction
`
`The multi-DSP and ASIC based acoustic front-end
`(AkuFE) has been developed in the framework of the
`SPICOS system/11 and is build up ofthree boards: Aso
`called master board and two slave boards. The system is
`designed for the analysis and synthesis of telephone
`quality speech (8KHz sampling frequency and 3.2 KHz
`bandwidth) and high quality speech (16 KHz sampling
`frequency and 6.4 KHz bandwidth). Figure 1 shows an
`general overview ofthe AkuFE system.
`
`Master·
`Board
`
`Slave·
`Boards
`
`Figure 1: The multi-DSP acoustic front-end system
`(AkuFE)
`
`The master board is the basic control system of an
`extendable multi-DSP architecture. The complete
`system is based on Texas Instruments TMS320C25
`digital signal processor VLSI chips and built as a
`VMEbus system. Within the distributed architecture of
`SPICOS /2/ it is plugged in a SUN workstation and
`employed for the real-time acoustic-phonetic decoding
`task. The master board of the AkuFE consists of an
`analog (A/D and D/A converter) and digital subsystem,
`while the slave board employs two TMS320 DSPs and a
`VLSI chip for VQ /3/. A local data and program memory
`is dedicated to each signal processor. Two slave boards
`can be connected to a master board, achieving a
`computational power of more than 100 MIPs. But the
`slave boards can be used also without a master board.
`
`The inter-DSPbus allows fast communication between
`the master and the slave processors without interfering
`with the VMEbus. A global memory scheme is realized
`in order to perform data transfers between the
`processors. Control is handled by interrupts. In order to
`test, download and control the system, the host
`machine has read and write access to the whole memory
`ofthe AkuFE system. Because ofits configurability and
`free programmability, the AkuFE system can be used
`in a wide range of signal processing applications, such
`as speech analysis and synthesis, speech coding and
`speech recognition. Used as a fast array processor for
`algorithms with extensive computational require(cid:173)
`ments, it can highly improve the performance of a
`standard workstation. The codebook generation /4/
`particularly can be accelerated by employing the VQ
`Chip on the slave board in connection with a host
`processor. Further applications are spectral analysis,
`instrumentation and image processing.
`
`A more detailed descri ption of the master board is gi ven
`in section 2. In section 3 the description of the slave
`DSP subsystems and the VQP follows. Finally an
`application of AkuFE system for real-time acoustic(cid:173)
`phonetic decoding is given in Chapter 4. Performance
`results for the specific application are given.
`
`It should be mentioned that the notation master and
`slave are not used here in the sense of VMEbus
`definitions.
`
`1 SPICOS (~iemens, ~hilips, !PO Continuous ~eech Recognition
`and U nderstanding) is carried out as a joint project and sponsored
`by the German Federal Ministry for Research and Technology
`(BMFT) under the grant No. ITM 8801 B9.
`
`EUROSPEECH '89, Paris, France, September 1989
`
`1586
`
`IPR2023-00035
`Apple EX1058 Page 1
`
`

`

`Sampie
`
`P1
`
`VMEbus
`
`Contr.-Reg.
`
`INTerrupt
`
`LEDs
`
`Ext. Buffer
`
`P2
`
`INTERwDSPbus
`
`Figure 2: Block diagram ofthe AkuFE master board
`
`2. The AkuFE Master System
`
`The master board 151 performs the initial signal
`preprocessing steps, like amplification, sampling, A/D
`and DIA conversion. Computationally
`intensive
`operations required for a spectral or time analysis can
`be executed on the DSP subsystem. Further task of this
`board is the control of data flow to the host machine
`over a so-called Mailbox. Figure 2 shows the general
`blockdiagram ofthe AkuFE master board.
`
`2.1. Analog Subsystem
`
`Two different audio channels, one for telephone and
`one for a high quality microphone input, are provided.
`The board allows the signal acquisition, filtering and
`sampling oftelephone quality speech (8KHz sampling
`frequency and 3.4 KHz bandwidth) and high quality
`speech (16 KHz sampling frequency and 6.8 KHz
`bandwidth). The corner frequency of the anti-aliasing
`filters for both sampling rates are software selectable.
`In order to increase the dynamic range ofthe 12 Bit A/D
`converter, a programmable gain preamplifier (0-24dB)
`is implemented.
`
`The DIAsubsystems consists of a 12 bit DIA converter
`and
`the associated interpolation filter with
`two
`different software selectable corner frequencies.
`
`The master board provides further outputs for
`telephone and a headset.
`
`a
`
`capable of executing ten million instructions per
`second, where most of the instructions require one
`cycle.
`A RAM based local data and program memory is
`dedicated to the DSP. The host processor has read and
`write access to the whole memory. A local data memory
`of 32 KWords and program memory of 16 KWords are
`provided on the master board. The remaining 32
`KWords of data memory are reserved for global
`memory communication areas of four
`further DSP
`subsystems located on two slave boards. The memory
`expansion and the fast communication to the slave
`board subsystems is performed via inter-DSP bus. Data
`exchange can be done via VMEbus interface as well.
`
`A data memory region of 1 KWord called MAILBOX is
`provided as dual-ported memory enabling a parallel
`access from both the DSP and the host processor
`without holding the DSP during a block data transfer
`routine. This allows, e.g.in case of signal acquisition, a
`real-time transfer of the preprocessed and analyzed
`data.
`
`For moreflexible VMEbus communication an interrupt
`control register is provided. Hold and reset control is
`performed by a further register which additionally can
`be used to toggle between the sampling rates. A third
`register allows the digital control ofthe gain amplifier.
`Several LEDs provide visual
`control of
`the
`communication.
`
`3. The AkuFE Slave System
`
`2.2. Digital Subsystem
`
`The second generation chip TMS320C25 from Texas
`Instruments provides a high performance 16 bitdigital
`subsystem for the AkuFE. The modified Harvard
`architecture of the processor allows a fast memory
`access and high data through-put with a 100 ns
`instruction cycle. The CMOS version of the DSP is
`
`The AkuFE slave is a powerful optional system that
`can also be used without the AkuFE master board 161.
`The two TMS320C25 based subsystems and the vector
`quantization processor (VQP) provides a computational
`power of about 50 MIPs and is appropriate for use as an
`array processor in many speech applications with high
`computational requirements. Two slave boards can be
`used in conjunction with the AkuFE master board.
`
`EUROSPEECH '89, Paris, France, September 1989
`
`1587
`
`IPR2023-00035
`Apple EX1058 Page 2
`
`

`

`Local memory is attached to each DSP subsystem. In
`order to test, download and control the system, the hast
`machine has read and write access to the whole memory
`of the AkuFE slave system. Four memory banks each
`with 1 Mbit are available, allowing storage of
`codebooks up to the size 8,000 in our application. The
`mapping of the VQP control register into the global
`memory area permits the control of codebook search
`from either one of the DSP subsystem of the AkuFE or
`from hast computer. Figure 3 shows the basic VQP-DSP
`interfacing.
`
`SANK 3
`
`SANK 2
`
`BANK 1
`
`BANK 0
`
`VECTOR QUANTIZATION
`PROCESSOR
`
`10MHz
`
`MICROPROCESSOR
`SYSTEM
`
`Figure 3: Vector Quantization Processor and DSP
`interfacing
`
`3.1. DSP Subsystems
`
`Two TMS320C25 based subsystems can be employed for
`parallel processing. Each subsystem has a dedicated
`da talprogram memory of 56 KW ords/64 KW ords. A
`data area of 8 KWords
`is reserved for global
`communication. Each global RAM area is accessible
`from all processors, including the VQPs. A priority
`controlled access is implemented. The master DSP has
`the highest priority. Theserialinput of each DSP can
`be used to build up a second communication path to
`other processors.
`
`The very highly flexible interrupt handling allows the
`implementation of various software architectures and
`communication schemes. A bus arbiter controls the
`access to the internal DSP bus.
`
`3.2. Vector Quantization Processor (VQP)
`
`As codebook search is. one of the most computationally
`extensive operations in speech processing, a VLSI Chip
`using high concurrency was developed to perform this
`task
`for
`real-time applications
`/3/. The VQP
`implements a full search algorithm. lt supports two
`software selectable distance measures for the search :
`Euclidian and city-block. The VQP is able to handle
`vectors of variable dimensions,
`codebooks
`of
`programmable size. Each codebook vector can be
`extended to up to 64 dimensions of 8 Bit components,
`without noticeable lass of performance. The chip
`delivers the codebook entry and the distance ofthebest
`codebook candidate.
`Finally, some important features ofthis ASIC should be
`mentioned:
`
`• on-chip cache RAM for input vector
`• vector dimensions u p to 64 wi th word-length of 8 bi t
`• 4 software selectable codebook banks, each oflMbit
`•
`throughput rate of 107 vector components/s
`• on-chip watchdog
`• 16 bitparallel interface
`easy communication with the hast processor
`•
`• needs two addresses in the host addressing space
`
`4. Real-time Acoustic Phonetic Decoding
`
`Within the SPICOS system, a man-machine dialogue
`system allowing data queries about office activities
`using a vocabulary of about 1,000 words, the AkuFE is
`used for real-time acoustic phonetic decoding and
`diphone synthesis /8/. The acoustic-phonetic decoding is
`based on articulatory features which describe the place
`the manner
`(labial, dental, alveolar, etc.) and
`(consonant, liquid, nasal, plosive, etc.) of articulation
`171 The feature vector generated from these categories
`is extracted every 10 ms and used by subsequent
`Hidden Markov Models for the recognition of subword
`units. It has been shown /7/ that the mapping of the
`speech signal onto articulatory categories has several
`advantages, such as the reduction ofthe feature set and
`speaker invariability. For the place of articulation,
`formants are required. Instead of implementing a
`computationally expensive root solving algorithm to
`calculate the LPC poles, a faster alternative approach
`using vector quantization was chosen. The spectral
`peaks in precomputed LPC spectra are stored in a
`codebook and used as formant candidates.
`The following processing steps are performed on the
`AkuFE: The master board performs the primarily
`signal
`preprocessing
`steps,
`like
`amplification,
`
`EUROSPEECH '89, Paris, France, September 1989
`
`1588
`
`IPR2023-00035
`Apple EX1058 Page 3
`
`

`

`/3/ E. Preiss, A. Stölzle, W. Drews and J. Pandel:
`"Architecture of a CMOS VLSI Vector Quantization
`Processor", Proc. of the EUSIPCO 88, Grenoble/
`France (1988), Vol. 3, pp.1241-1244
`/4/ T.M. Liu and H. Höge: "Phonetically Based LPC
`Vector Quantization ofHigh Quality Speech", Proc.
`of this Conference
`/5/ G. Gohn and A. Peham: "Pflichtenheft, Akustisches(cid:173)
`Frontend", Interna! Report, Siemens Wien (1986)
`"AkuFE-Slave, Spezifikationen",
`/6/ S.
`Szikora:
`Interna! Report, Siemens Wien (1988)
`171 0. Schmidbauer : "Robust Modelling of Systematic
`Variabilities in Continuous Speech Incorporating
`Acoustic Articulatory Relations", Proc. ofte ICASSP
`'89, Glasgow/U.K. (1989), pp 616-619
`/8/ A. Aktas and H. Höge: "Real-Time Recognition of
`Subword Units on a Hybrid Multi-DSP/ASIC Based
`Acoustic Front-End", Proc. of the ICASSP '89,
`Glasgow/U.K. (1989), pp. 100-104
`/9/J. Sokat : "Untersuchung der Autokorrelation(cid:173)
`Ladder-Algorithmen zur Berechnung von Prädik(cid:173)
`tionsfilter-Koeffizienten
`in
`Integer-arithmetik",
`Diplom-Thesis, Uni Duisbu"rg/W. Germany(1987)
`
`sampling, AfD and DIA conversion with a linear
`resolution of12 bits. The analogue speech signal is low(cid:173)
`pass filtered at a corner frequency of 6.4 kHz and
`digitized with a sampling rate of 16 KHz. Further
`tasks of this board are the control of data flow over a so(cid:173)
`called Mailbox and the computation of some basic
`parameters like signal energy and zero crossings.
`The acoustic analysis and synthesis algorithms are
`implemented on the second board. In order to reach
`high spectral resolution and detect reliable formants in
`the 0-6.5 KHz region of the spectrum, a 16th order LPC
`analysis is performed. The LeRoux and Guegen
`algorithm
`is
`implemented
`in mixed 16/32-bit
`arithmetic /9/.
`
`The VLSI VQ-Chip performs a full search over 4,000
`LPC codebook entries represented by 16 coded
`reflection coefficients each coded with 8 bits. The
`codebook entries deliver formant candidates and
`corresponding sub-band energies for the succeeding
`formant tracking algorithm. Both formant tracking
`and articulatory analysis with the composition of the
`AFV consisting of the occurrence probabilities of
`articulatory categories /7/ are performed on the host
`machine.
`The homogeneaus articulatory feature vector which is
`composed of probabilities describing the manner and
`place of articulation, serves as an input parameter for a
`phoneme recognition scheme based on HMMs. Finally
`the phoneme models are concatenated in order to model
`400 subword units of consonant clusters and syllabic
`these
`nuclei. The Viterbi algorithm to recognize
`clusters is now going to be implemented on an
`accelerator board with an ASIC. This board will be
`built up of a high performance 32 bit floating-point DSP
`and and a RISC processor.
`
`5. Performance
`
`The processing steps, from the sampled speech data
`through the calculation of the reflection coefficients,
`require about 8 ms on a TMS320C25, leading to a frame
`rate of 10 ms. Besides the weighting and calculation of
`the autocorrelation coefficients, which require about
`1.5 ms, about 0.5 ms is necessary for the double(cid:173)
`buffered processing and real-time signal acquisition. As
`the VQ processor can perform a full search over the
`total codebook in Iess then 9 ms, the processing can be
`performed in real-time.
`
`6. References
`
`111M. Brenner, H. Höge, E. Marschall and J. Romano:
`"Word Recognition in Continuous Speech Using a
`Phonological-Based Two Network Matehing Parser
`and a Synthesis Based Prediction", Proc. of the
`ICASSP '89, Glasgow/U.K. (1989), pp. 457-460
`/2/ H. Höge : ''Verteilte Prozessarchitektur für ein
`sprach verstehendes Ech tzei tsystem", ·. Proc. of the
`DAGM-Symposium, Braunschweig!W. Germany
`(1987), pp.l06-110
`
`EUROSPEECH '89, Paris, France, September 1989
`
`1589
`
`IPR2023-00035
`Apple EX1058 Page 4
`
`