ISCA Archive
`http://www.isca-speech.org/archive
`
`2nd European Conference on
`Speech Communication and Technology
`EUROSPEECH ’91
`Genova, Italy, September 24-26,1991
`
`SPEAKER INDEPENDENT CONTINUOUS HMM-BASED RECOGNITION
`OF ISOLATED WORDS ON A REAL-TIME MULTI-DSP SYSTEM
`
`Abdulmesih Aktas and Klaus Zünkler
`
`Siemens AG, Corporate Research & Development
`Otto-Hahn-Ring 6,8000 München 83, Germany
`
`Abstract
`This paper describes a speaker independent
`continuous Hidden Markov Model recognizer
`implemented on a real-time multi-DSP system.
`Training and recognition are based on continuous
`mixture density HMMs for phonemes. Context
`dependent triphone models are used and the Viterbi
`algorithm is applied for both training and recognition.
`The system is implemented on a workstation with an
`integrated multi-DSP based acoustic front-end
`employing three Texas Instruments TMS320C25 signal
`processors and a Siemens ASIC for vector quantization.
`In spite of the simplifications made in order to reduce
`the high computational requirements for the
`continuous mixture densities, the system has a
`recognition rate of 99.5% for the speaker independent
`German digit task with telephone quality.
`
`Keywords: Speech Recognition, Hidden Markov Model,
`mixture density, speaker independent, real-time.
`
`1. Introduction
`
`Hidden Markov Models (HMM) with continuous
`mixture densities as stochastic models for speech
`recognition have been shown to be well-suited for
`modeling spectral and temporal variabilities in speech
`[4], The computation of the emission probabilities re­
`quired for the probability density functions is usually a
`highly time consuming task. Therefore most real-time
`applications of HMM-based recognizers on digital
`signal processors (DSP) use discrete models (e.g. [3]).
`For the implementation of the continuous mixture
`density HMMs powerful floating point DSPs or special
`ASICs have to be employed. We use in our approach a
`vector quantization VLSI-chip (VQ) for producing the
`emission probabilities. For the feature extraction
`integer DSPs are employed. The Viterbi search is
`implemented on a RISC-based workstation.
`The recognizer was evaluated on two different
`speaker independent German digit databases, whereby
`recognition rates of 99.5% were achieved. More details
`on the recognition experiments are given in [6,7].
`
`The following section of our paper describes the
`structure of our models. Section 3 deals with the real­
`time multi-DSP front-end, while section 4 presents
`implementation aspects of the recognition algorithm on
`the hardware. The VQ realization of the emission
`probability computation is explained in more detail.
`Finally some results conclude the paper.
`
`2. Model Structure
`
`As experiments with word models have shown a
`poor ability of discriminating confusable words,
`phoneme models are used in our system [6]. The
`phonemes are modelled with context dependency
`because their acoustic realizations are often heavily
`affected by coarticulation. The HMMs that we use for
`each phoneme consist of three segments for modeling
`the beginning, the middle and the end. Each segment
`consists of two states with tied emission probabilities.
`The first and the third segment consider coarticulatory
`effects due to the transitions to the neighboring
`phonemes, whereas the second segment stands for the
`middle part which is less affected by the phonetic
`context.
`According to a pronunciation lexicon word models
`are concatenated of phoneme models. During
`recognition only phoneme sequences contained in the
`lexicon are used. On the segment layer triphone
`models, shown in Figure 1, are the basis. In the example
`the concatenation of three phonemes (/d/, /r/ and /ai/) to
`the German word 'drei' is given. The three phonemes
`are embedded into silence models (/si/). The second line
`shows the context dependent segments (si, sld, d, dr, dr
`and so on) with their associated indices, which indicates
`the adjacent or predecessor segment of the current
`phoneme for this word. This kind of modeling allows
`robust and high recognition performance.
`On the segment layer the possible . number of
`triphones is reduced because every middle segment of a
`phoneme model is adjacent only to segments belonging
`to the same phoneme. So the middle segments of the
`phoneme models do not have any context information.
`Each beginning segment of a phoneme is succeeded by
`the same context independent middle segment. Thus
`
`EUROSPEECH ’91,Genova, Italy, September 1991
`
`1345
`
`IPR2023-00034
`Apple EX1059 Page 1
`
`

`

`phonetic transcription:
`/d/
`/si/ 1
`1
`context debendent segmentd:
`si । s’d i d । dr । dr
`i
`i
`i
`i
`
`/r/
`
`1
`i r
`i
`
`/ai/
`r • ।
`»
`•
`ai । ai । ar
`* si
`i
`_i
`evo
`
`/si/
`si
`
`Figure 1: Context dependent phoneme
`based word model of the
`German digit 'drei' (three).
`
`---------------1......... 1
`1
`dual-ported memory enabling a parallel access from
`the triphones are reduced into a mixture of bigrams and
`both the DSP and the host processor. This allows in case
`context independent segments. With an inventory of 40
`of signal acquisition, a real-time transfer of the
`phonemes approximately 3000 segment models would
`preprocessed and analyzed data. The inter-DSPbus
`be possible. In practice, with a small digit vocabulary
`allows fast communication between the signal
`this is reduced to 90 segment models.
`processors on the master and the slave boards without
`interfering with the VMEbus. A global memory scheme
`is realized in order to perform data transfers between
`the processors. For a flexible VMEbus and inter-DSP
`communication an interrupt control is provided [1].
`The VQ processor is able to handle vectors of
`variable dimensions and codebooks of programmable
`size. Each codebook vector can be extended up to 64
`dimensions of 8 bit components, without noticeable loss
`of time performance. Four software selectable codebook
`memory banks each with 1 Mbit are dedicated to the
`processor. The chip can be controlled from either one of
`the DSP subsystems of the AkuFE or from the host. The
`on-chip cache RAM of the VQ processor allows parallel
`load of a vector during the search. Within 1 second the
`VQ can compare 107 vector components of the codebook
`to a given vector. Two different distance measures
`(City-block and Euclid) are supported. The chip delivers
`the codebook entry and the distance of the best
`codebook candidate.
`
`3. Multi-DSP Acoustic Front-End
`
`The recognition system is implemented on a
`UNIX workstation employing a multi-DSP based
`acoustic front-end (AkuFE) [1] which was developed
`within the speaker adaptive continuous speech under­
`standing system SPICOS [4]. In the SPICOS system the
`AkuFE is employed with the acoustic-phonetic
`decoding and the real-time recognition of subword units
`[2], The AkuFE is designed for real-time speech
`processing tasks and is a configurable VMEbus system
`employing Texas Instruments TMS320C25 signal
`processors and a Siemens ASIC for vector quantization
`(VQ). A powerful system can be configured using three
`boards (one master and two slave boards) with a total
`computational power of more than 100 MLPs. In this
`case five signal processors and two VQ processors can
`work in parallel. Figure 2 shows a general
`configuration of an extended system.
`
`4. Real-Time Implementation
`
`Real-Time Considerations
`'
`Training and recognition in our system are based
`on continuous mixture density HMMs for phonemes as
`used in the SPICOS system [4], The main difference
`besides recording conditions and sampling frequencies
`is that context dependent models are used. The Viterbi
`algorithm is implemented for both, training and
`recognition. Our models are left to right first order
`Markov models with a fixed A matrix, N-dimensional
`feature vector and a continuous mixture density
`function p(x) with M modes, which is based on Gaussian
`density functions:
`M
`
`P(Xt) =
`
`m=l
`
`where xt represents the input frame vector, cm is the
`mixture weight for the m-th mode, Um is the mean
`
`Figure 2: The multi-DSP acoustic front-end system
`
`The master board is the basic control system of an
`extendable multi-DSP architecture. It consists of an
`analogue (A/D and D/A converter) and digital, Texas
`Instruments TMS320C25-based subsystem, whereas
`the slave board employs two TMS320C25 DSPs and a
`VLSI-chip for VQ. A local data and program memory is
`dedicated to each signal processor. A data memory
`region of 1 KWord is provided on the master board as
`
`EUROSPEECH ’91,Genova, Italy, September 1991
`
`1346
`
`IPR2023-00034
`Apple EX1059 Page 2
`
`

`

`vector of the mixture mode m and Cm is the covariance
`matrix.
`In order to achieve real-time processing some
`assumptions concerning the emission probabilities are
`made [6]:
`.
`The covariance matrix C is assumed to be
`•
`diagonal.
`It is assumed that the contribution of the nearest
`mixture density is much larger than the con­
`tribution of the others, thus the probability is
`computed only for -the best fitting mode k for a
`given input vector Xf Additionally the variances
`of all mixture densities are considered to be equal.
`This results in the reduction of formula (1) to:
`(2)
`Pk(xt) =
`Ai
`where Ki and K2 are constants which depend on the
`variance of the distributions. Simulation experiments
`have shown that this is no limitation if there are
`enough mixtures available for modeling the density
`functions. With these simplifications the emission
`probability depends only on the mean of the best fitting
`mixture density and its mixture weight. In the
`logarithmic domain, the computation of the probability
`simplifies to an Euclidian distance
`N
`- In pk(xt) = Ky - In cfc + K2 (*'" “
`n=l
`which can be easily computed by the special hardware.
`With the assumption of Laplacian density
`functions the mixture computation can be further
`reduced to the City-block distance and modified to a
`distance measure dkt derived from the probabilities:
`- In pi,^) Ki
`-----r-----------t
`A 2
`A2
`N
`1
`in Ck . .
`~
`= Q + > y Pin “ /¿kn| , With Ck =--- —
`(4)
`2
`n—1
`Ck is here a mode dependent constant.
`Previous experiments with both distance
`measures have not revealed significant differences in
`the recognition results.
`
`(3)
`
`•
`
`=
`
`The Viterbi search algorithm is split into two
`parts which run on different processors. The production
`of the emission probabilities for each frame is
`performed on the VQ processor while the calculation of
`the accumulated word model probabilities is processed
`on the host machine. Figure 3 shows the organization of
`the system moduls on the hardware.
`In the current system three DSPs and one VQ
`processor are employed for the preprocessing, feature
`extraction and emission probability calculation. The
`DSP on the master board and the VQ control processor
`work in multitasking modes. The first DSP handles
`
`Figure 3: The organization of the recognition moduls on
`the hardware.
`
`additionally the double-buffered organization of the
`real-time signal acquisition. The host machine receives
`for each 10 ms an interrupt which initiates a buffer
`read on the mailbox of the master board to get the
`distances for the current frame.
`
`Feature Extraction
`Two boards of the AkuFE system are employed for
`the recognition task. The first board of the AkuFE
`system performs the preprocessing steps, amplification,
`low-pass filtering at a corner frequency, of 3.4 kHz,
`sampling with 8 kHz and A/D conversion (12 bit
`resolution). After the preemphasis a 256-point
`Hamming window is multiplied to the speech frame
`with a step size of 10 ms. Further tasks of this board are
`the control of the slave DSPs and of the data flow over
`the Mailbox and global memory.
`Mel-scaled filter bank energies and their
`differences are used as feature vector. The time
`consuming computation of this vector is implemented
`in integer arithmetic on the slave board employing both
`of the DSP subsystems. A 256-point FFT is applied to
`the windowed signal. The spectrum is passed to a filter
`bank. The Mel-filter bank basically performs a
`summation of the FFT-channel energies within the
`selected frequency ranges varying according to the Mel­
`scale. The filter bank energies are then transformed to
`the log-domain. This allows a coding of each dimension
`with 8 bit resolution. In order to eliminate irrelevant
`energy variations for neighbored frequencies which are
`caused by the pitch, a cepstral smoothing is applied to
`the 24-dimensional spectras. The smoothing is done by
`a convolution of a liftering window transformed into the
`frequency domain. Additionally the total frame energy,
`the first and the second derivative of the Mel-vector is
`computed. After condensation of the derivates the
`feature vector has 52 dimensions.
`.
`
`Computation of the emission probabilities
`The feature vector, which is computed in real-time
`every 10 ms, is passed to the VQ stage in order to
`compute the local emission probabilities for the current
`
`EUROSPEECH ’91,Genova, Italy, September 1991
`
`1347
`
`IPR2023-00034
`Apple EX1059 Page 3
`
`

`

`frame. The organization of the codebook is as follows:
`For each segment there are about ten codebook entries
`containing the mean vector and the mixture weight of
`each density. The VQ processor performs in real-time a
`search over all codebook segments. As each segment is
`assigned to a subword model the vector quantizer chip
`computes the distance to the mean of the best fitting
`mixture density. This is based on the simplifications
`given above in this section (equation (4)). The minimum
`distance for each frame and segment model is passed to
`the workstation which evaluates the word-HMMs in
`parallel and in real-time. As the local emission
`probabilities are represented in the log-domain, the
`calculation of the word probabilities is reduced to a
`.
`summation.
`The VQ processor is controlled by one of the slave
`DSPs. After each search within a segment the VQ
`generates an interrupt, causing a distance and index
`read by the DSP.
`
`Viterbi Search
`As our system is a phoneme based recognition
`system each word in the lexicon is built up from a
`sequence of phoneme units. For the purpose of word
`boundary detection a silence model is added to the
`beginning and end of each model. The number of states
`in the word models defines the search space for each
`word. This remains static for each vocabulary. In our
`case of small vocabularies the search space for all words
`is prepared in advance. For the digit task the maximum
`search space includes 256 states (sum of all states in the
`word models). For each search space the word model
`probability has to be calculated. The computation can
`be done in a parallel manner by considering the current
`feature vector in each search space of each word. This
`requires about 8 bytes memory per state. A 10-MIPs
`workstation is necessary in order to handle this task in
`real-time.
`The end detection of the utterances is checked
`automatically during the search. The best word
`candidate is evaluated at the final pause state of each
`word model and if a candidate is the winner for at least
`300 ms it is selected as the best word.
`The multi-DSP system is also well-suited for the
`training purposes of the Hidden Markov Models
`because the computation of the indices of the best
`modes by the VQ processor allows us to reestimate the
`mean values of the distributions.
`
`5. Results and Conclusion
`
`The recognition experiments for this work are
`described in [6], The German digit database we use
`contains 11 words (0-9 and and an alternative
`pronounciation for the German digit 'two'(zwo)). The
`training set consisted of utterances spoken by 15 male
`and 15 female speaker each with four versions in
`
`average. Speech data of another 30 male and female
`speakers (440 utterances) was used for test experi­
`ments. The recording was performed over an ISDN
`telephone device with a sampling rate of 8 kHz. The
`recognition rate is 99.5% for the test database.
`In the current demonstration system the Mel-
`vector calculation requires about 20 MIPs. The next
`generation DSP (e.g. TMS320C50) will allow us to
`perform the whole Mel-vector computation on one
`DSP.With a faster FFT, which runs now on the first
`slave DSP it is possible to reorganize the moduls and
`extend the number of the segments. That means, new
`words can be added to the lexicon. Currently the VQ
`requires 6 ms and about 40 MIPs for the search, given a
`segment based codebook. A possibility to double the
`vocabulary is to employ two VQ processor at a time. The
`time intensive implementation of the feature extraction
`algorithms is justified under the aspects of transferring
`the algorithms to VLSI chips. Expensive silicon space
`can be saved on this way.
`As our system is phoneme based the emission
`probability production can be easily extended to
`continuous speech recognition applications. For the
`1000 word vocabulary of SPICOS around 130 context
`dependent segments has to be considered in real-time.
`
`References
`
`[1] A. Aktas and H. Hoge: ”Multi-DSP and VQ-ASIC
`Based Acoustic Front-End for Real-Time Speech
`Processing Tasks”, Proc. EUROSPEECH '89, Paris,
`pp.586-589
`[2] A. Aktas and H. Hoge: "Real-Time Recognition of
`Subword Units on a Hybrid Multi-DSP/ASIC Based
`Acoustic Front-End”, Proc, of the ICASSP '89,
`Glasgow, pp.101-103
`[3] A. Ciaremella et.al.: ”A PC-housed speaker
`independent large vocabulary continuous telephonic
`speech recognizer”, in these Proceedings
`[4] H. Hoge: "SPICOS H - A speech understanding
`dialogue system”, Proc, of the Int. Conf, on Spoken
`Language Processing (ICSLP), Kobe, Japan (1990)
`[5] H. Ney and A. Noll: "Phoneme modeling using
`continuous mixture densities”, Proc, of the ICASSP
`'88, New York, pp.437-440
`[6] K. Zunkler: "An ISDN speech server based on
`speaker independent continuous Hidden Markov
`Models”, Proc, of NATO-ASI, Cetraro, July 1990
`[7] K. Zunkler: "A Discriminative Recognizer for
`Isolated and Continuous Speech Recognition Using
`Statistical Information Measures”, in these
`Proceedings
`
`EUROSPEECH ’91,Genova, Italy, September 1991
`
`1348
`
`IPR2023-00034
`Apple EX1059 Page 4
`
`