We describe the construction and use of a laboratory facility for testing the performance of speech-to-speech translation devices. Approximately 1500 English phrases from various military domains were recorded as spoken by each of 30 male and 12 female English speakers with variation in speaker accent, for a total of approximately 60,000 phrases available for experimentation. We describe an initial experiment using the facility which shows the impact of environmental noise and speaker variability on phrase recognition accuracy for two commercially available one-way speech-to-speech translation devices configured for English-to-Arabic.

The intelligibility of speech transmitted through low-rate coders is severely degraded when high levels of acoustic noise are present in the acoustic environment. Recent advances in nonacoustic sensors, including microwave radar, skin vibration, and bone conduction sensors, provide the exciting possibility of both glottal excitation and, more generally, vocal tract measurements that are relatively immune to acoustic disturbances and can supplement the acoustic speech waveform. We are currently investigating methods of combining the output of these sensors for use in low-rate encoding according to their capability in representing specific speech characteristics in different frequency bands. Nonacoustic sensors have the ability to reveal certain speech attributes lost in the noisy acoustic signal; for example, low-energy consonant voice bars, nasality, and glottalized excitation. By fusing nonacoustic low-frequency and pitch content with acoustic-microphone content, we have achieved significant intelligibility performance gains using the DRT across a variety of environments over the government standard 2400-bps MELPe coder. By fusing quantized high-band 4-to-8-kHz speech, requiring only an additional 116 bps, we obtain further DRT performance gains by exploiting the ear's insensitivity to fine spectral detail in this frequency region.

Users of low data rate voice coders now demand good speaker recognition in addition to quality and intelligibility. Speaker recognizability was one of the selection criteria for the DoD Digital Voice Processor Consortium in selecting a new standard 2400 bps algorithm. We have developed a speaker recognition test based on same-different judgments for pairs of sentences spoken by 10 male and 10 female speakers. Preliminary experiments showed that the sensitivity of the test is similar to that of currently used acceptability and intelligibility tests. Male and female talkers produced somewhat different rank orderings, but averaging over males/females gave results that make sense when compared with intelligibility and acceptability scores for the same coders. The results of these experiments determined the speaker recognition test procedures that were used in the 2400 bps coder final selection process.

The U.S. Department of Defense's Digital Voice Processing Consortium (DDVPC) is in the process of choosing a new 2400 bps vocoder standard. As part of the selection process, a study was conducted to determine the test measures to be used for the evaluation of the intelligibility and quality of competing vocoder algorithms. The DDVPC requirements for a vocoder differ substantially from commercial interests. Severe ambient acoustic noise backgrounds coupled with hostile transmission channel conditions must be considered. The commercially popular mean opinion score (MOS) test method was evaluated as a replacement or companion test of quality to the diagnostic acceptability measure (DAM), traditionally used by the DDVPC. The study showed that a properly structured MOS test can achieve equal resolution, reliability, and validity to that of the DAM at equivalent costs. Any MOS test series must be structured to minimize contextual effects. Certain severe conditions must use the degraded mean opinion score (DMOS) test method to achieve usable resolution.

The evaluation of the intelligibility of voice communication systems required to operate over severely disturbed transmission channels is not possible with standard test methods that incorporate isolated words presented at a set rate or pace. With advancements in the quality and intelligibility of narrowband digital voice communications systems has come increased interest in the application of these systems to a wide number of areas. A unique modification to the diagnostic rhyme test (DRT) method that provides for the evaluation of intelligibility for these systems is presented. A controlled experiment tests the validity of the modified method with an experienced crew of DRT listeners. The results of this experiment show that a modification of the DRT for interrupted systems provides valid measures of intelligibility. Also discussed are the modifications required for scoring an interrupted DRT and the software developed to provide scoring, storage, and comparative analysis of the results.>

The Canonical Coordinate method of signal compression offers the possibility of incorporating psychoacoustically derived non-euclidean error metrics into speech compression, but requires highly parallel matrix computations. We describe an actual implementation of Canonical Coordinate processing on a systolic array of 64 computational nodes connected as a ring. We discuss microcoding support software developed for the array, and systolic algorithms for discrete Fourier transform, matrix-by-vector multiplication, and sorting. This implementation has been used to develop design parameters for a self-contained real-time systolic architecture.

This paper describes the implementation of a matrix array processor oriented technique that can transform time sampled speech waveform data, by way of the frequency domain, into a psendo-canonical coordinate domain whose components can be ordered in such a manner as to commit the smallest signal reconstruction error, relative to a given non-euclidean quadratic error criterion, when certain ordered components are not used in the reconstruction process. The error metric can be specified, in either the time or frequency domain, in the form of a complex valued hermitian matrix, thus allowing a very large number of degrees of freedom that can be used to incorporate certain statistical, dynamical and psychoacoustic attributes of the speech data into the analysis and reconstruction process. For the case of a simple euclidean error metric, the pseudo-canonical coordinate process is shown to reduce to the well known principal component, or Loevé-Karhunen, signal representation technique.

Speech level measurements were made for a collection of Diagnostic Rhyme Test (DRT) word lists as used for intelligibility testing of voice communications systems. These measurements were used in the generation of a subsequent set of DRT source tapes for which all DRT list segments on the tapes were equalized with respect to speech level. The effects of this equalization on intelligibility scores was evaluated for unprocessed and processed speech (both parametric and waveform coding algorithms). The manipulation of the speech levels for the DRT lists did not result in significant, predictable, and replicable effects on either system mean scores or on the DRT score variability across lists. The absence of effects was noted for all speech processing conditions and for both male and female speakers. Practical advantages in the use of the equalized word lists are discussed.

This paper describes the use of a canonical signal compression and modelling technique that permits the minimization of certain non-euclidean types of signal resynthesis error criteria. The technique is based on the construction of a non-orthogonal transformation from the original sampled signal representation, in either the time, frequency or spatial domain, to a special canonical coordinate domain. The parameters characterizing this transformation are then chosen to minimize the specified error criterion, for each level of truncation of the canonical coordinate based signal representation. A mechanism for factoring this canonical coordinate transformation into an eigenvector-eigenvalue based correlation simplification process and an error metric simplification process, is described. This mechanism is shown to involve principal component (or Loeve-Karhunen) analysis as an intermediate step in the complete canonical coordinate determination process, and can lead to a substantial simplification in the computational complexity that is entailed in handling a class of non-euclidean error criteria.