This paper describes the recognition of container code characters in the project Mocont-II, where container images are taken in largely varying light situations. The recognition system has to deal with gray-level characters showing a wide variability of brightness and contrast, varying inclination, segmentation uncertainties, damaged characters and the presence of shadows. Different sets of features were extracted directly from gray-level images, and a minimum distance classifier with a weighted metric was used for recognition. To achieve good recognition performances, the feature weights and the prototype sets were optimized by a new gradient-based learning algorithm that maximizes a fuzzy recognition rate functional. 1.

In segmentation-by-recognition methods, words are recognised by searching for the aggregation of basic segments (strokes) that best matches one of the sequences of letters allowed by the lexicon. The presence of spurious strokes ("extra ink") impairs recognition, since they have to be associated anyway with some letter of the word. The paper describes how we solved this problem by properly defining initial crisp evaluations to filter out unambiguous extra ink strokes, and then, for ambiguous strokes, by defining a fuzzy measure that evaluates the plausibility with which they can be associated to a class of extra ink instead of to a letter class, and finally by devising search algorithms that allow extra ink strokes to be "consumed" without associating them to letters: it is the matching algorithm that decides if it is less costly to consider a hypothesis as extra ink or to associate it to some letter.

The paper addresses the problem of removing unwanted variations in cursive handwriting by estimating and compensating reference lines and slant angle. To this purpose, the paper introduces the mathematical framework of generalized projections (GP), an extension of the well known projection profiles (PP). This framework proved to be quite effective and robust in the estimation of those parameters. The particular application of GP to skew angle estimation, slant estimation and reference lines detection is then described. Experimental results are finally presented and discussed.

In this paper we describe two novel algorithms for variable frame analysis and word matching to be used in connected word recognition. The main characteristic of the algorithm for variable frame rate coding is that of smoothing out the non significant spectral fluctuations in speech. To take the different frame durations into proper account, the standard DP-matching algorithm was modified. The proposed modifications serve the purpose of favouring the alignment of similar phonetic events (similar also according to duration) and of emphasizing to the right extent the importance of transient sounds. Experiments on speaker-trained connected digit recognition have displayed higher correct recognition scores than the standard recognizer (10 msec. frame rate), even with a computational load reduced to about 1/10 of the original.

Almost all CSR systems presently in practical use are based on whole-word template matching. Although their performances are quite high, a few problems arise due to the use of whole words as ba sic units. They are related to storage require ments, coarticulation effects at the junction be tween words, discrimination of similar words and training time, especially for large vocabularies. In this paper we report a very positive experi ence which is being made with a system based on very short sub-word units, called "diphones". With the approach described in this paper, the above mentioned problems are much alleviated, with no penalty on performance. The first part is devoted to the presentation and discussion of the peculiar characteristics of the diphones, of the language model based on them and of the overall recognition system. Then a set of procedures used for training the system on a new application and for extracting the diphone templates for any new speaker are briefly de scribed. Finally we report and discuss the results of different tests performed on various recogni tion tasks.

This paper describes a connected speech recognition system based on Markov models. The performance of this system was analyzed and compared with that of a system which uses prototypes of words instead of Markov models. Some preliminary results are reported with reference to the recognition of connected digits.

The spectrum of additive noise was systematically varied according to three variables: correspondance between the noise and speech spectrum (4 levels), equal weighting or weighting according to the width of the Articulation Index bands, SNR of the output signal (2 levels). All combinations of these 16 noise conditions were added to high quality speech sampled at 8kHz, and presented to listeners for judgment. Difference judgments were analyzed by a multidimensional scaling procedure to gain information about the number and nature of the important subjective variables. Overall quality ratings were obtained as a measure of the improvement in perceived quality due to the spectral shaping of the additive noise. Several objective measures were also tested for their sensitivity in predicting the overall quality ratings.

The paper describes a preliminary implementation of a continuous speech recognition system based on the concept of diphone spotting. This consists of continuously measuring the similarity of the current portion of signal pattern with a complete set of selected phonetic events, called diphones because the most significant of them are transitions between pairs of phonemes. The entire set of measures feeds a linguistic decoder that operates on a state space representation of the language, whose states are the diphones that compose the words of the lexicon. Durational constraints and optional phonological rules are included in the language representation. The linguistic decoder recognizes the sentence as that path through the network which attains the highest cumulative similarity score. Additionally, the decoder has the task of detecting the end of the sentence. A preliminary test on 50 sequences of 3 to 7 connected digits gave recognition rates as high as 99.6% on digits, or 98% on sequences, thus confirming the validity of this approach.

In this paper we describe a multiprocessor architecture used in a real-time interactive system based on the recognition of sentences composed of sequences of isolated utterances. Anyway, the same general criteria could as well be used in the continuous speech framework. Particular attention is devoted to the control machanisms which supervise the system and, finally, detailed examples are given.

Objective quality measures provide an economic and practical alternative to tedious and expensive subjective tests. Furthermore, some recently proposed measures predict quality values which correlate quite well with the subjective ones. The objective measures are particularly useful to test the influence of design parameters during the laboratory development of codec algorithms. In this paper the architecture of a measuring equipment to conduct objective quality measures on speech waveform coders is presented and discussed. Only the input and output signals from the coder are needed and an identification algorithm is used to separate the distortion effects introduced by the coder from other effects such as gains, time shifts, low-pass filtering, and so on. The choice of the most appropriate kind of signal for the identification and measurement steps is a problem also investigated. Particularly, results concerning mathematically defined speech-like signals are reported. The performance of the method is investigated under different conditions and the correlations with subjectively determined results are also reported for a number of ADPCM coders with widely varying parameters.

Objective quality measures provide an economic and practic alternative to the tedious and expensive subjective tests. Furthermore some recently proposed measures predict quality values which correlate quite well with the subjective ones. In this paper the architecture of a measuring equipment to conduct objective quality measures on speech waveform coders is presented and discussed. Only the input and output signals from the coder are needed and an identification algorithm is used to separate the distortion effects introduced by the coder from other nondistorting effects as gains, time shifts, low-pass filtering and so on. The performance of the method is investigated under different conditions and the correlation with subjectively determined results is also reported for a number of ADPCM coders with largely varying parameters.

Spectral shaping of quantizing noise is an important issue for achieving "transparent" speech quality with low bit rate wave-form coding. To obtain optimum quality in an Adaptive Predictive Coder with Noise Feedback (APC-NF), we have defined a class of functional forms for the noise spectrum, controlled by three parameters, that could span as completely as possible the space of reasonable shapes it could assume. The shaping functions are composed of two factors: a time-varying one, related to the speech spectrum, which accounts for the masking mechanism, and a time-invariant one which accounts for the critical bands phenomenon in the human ear. This formulation led also to a filter design computationally elegant and stable. Subjective tests have shown that optimum quality is achieved when the variable term in the noise spectrum duplicates the square root of the speech power spectrum, while the fixed term consists of a slight pre-enphasis.

An experiment has been performed to study the perceptual characteristics of speech processed by ADPCM. We created 18 three-bit and four-bit coders spanning a wide range of quantizer adaptation parameters. Subjects judged the difference between each pair of coders and rated the quality of each coder individually. The difference data reveal three important perceptual dimensions (overall clarity, signal vs. background distortion, muffled vs. hoarse) which are related to various objective measures of coder performance. Overall subjective quality is well predicted by segmental SNR and even better by a linear combination of measures of granular distortion and overload distortion.

Most of the DPCM systems for speech transmission, employ a digital transversal predictor. In such a way an all-pole vocal tract model is assumed; this is not always good either for a fixed or time-varying model. To overcome this drawback some recursive-filter predictors (e.g., Kalman predictors) have been proposed, but not examined for real speech. This paper develops one of these procedures, and, by exploiting an average state-space model of the vocal tract, analyzes the system performance in the case of real voice.