Neural text-to-speech (TTS) systems have recently demonstrated the ability to synthesize high-quality natural speech. However, the inference latency and real-time factor (RTF) of such systems are still too high for deployment on devices without specialized hardware. In this paper, we describe StreamSpeech – an optimized architecture of a complete TTS system that produces high-quality speech and runs faster than real time with imperceptible latency on resource-constrained devices by utilizing a single CPU core. We divide the standard TTS processing pipeline into three phases with respect to their operating resolution and optimize them separately. Our main novel contribution is the introduction of a lightweight convolutional acoustic model decoder, which enables streaming and low-latency speech generation. Experiments show that the resulting complete TTS system achieves 79 ms latency, 0.155 RTF on a low-power notebook x86 CPU and 276 ms latency, 0.289 RTF on a mid-range mobile ARM CPU with no noticeable difference in the quality of the generated speech.

We present a new general and language independent approach to the noisy text correction problem developed and implemented in the framework of the CULTURA project. We briefly describe the core candidate generator, REBELS, the complete system concept, its efficient implementation based on functional automata and its immediate applications. The quality of the whole system is empirically established in different experimental settings where language and noise sources are varied.

We describe a novel approach for the extraction of spelling variations from a list of instances. It relates distinctive infixes to distinctive infixes of referenced words. The distinctive infixes are extracted automatically from a (multi)set of instances and a referenced dictionary without any additional expert knowledge. Based on the spelling variations retrieved during a learning(training) phase we develop a correction algorithm which suggests and ranks candidates for a particular noisy word. The main advantage of our approach is that it provides good corrections for the unobserved noisy words while it is almost perfect on words observed during the learning. Our experimental results of the normalisation of a typical reference corpus of Early Modern English letters, [1], significantly improve over previous results of VARD2, [2]. We also achieve better results than those reported in [3] and [4] on the OCR-correction of the TREC-5 Confusion Track corpus,[5].

Lexical text correction relies on a central step where approximate search in a dictionary is used to select the best correction suggestions for an ill-formed input token. In previous work we introduced the concept of a universal Levenshtein automaton and showed how to use these automata for efficiently selecting from a dictionary all entries within a fixed Levenshtein distance to the garbled input word. In this paper we look at refinements of the basic Levenshtein distance that yield more sensible notions of similarity in distinct text correction applications, e.g. OCR. We show that the concept of a universal Levenshtein automaton can be adapted to these refinements. In this way we obtain a method for selecting correction candidates which is very efficient, at the same time selecting small candidate sets with high recall.

Abstract Problems in the area of text and document processing can often be described astext rewriting tasks: given an input text, produce a new text by applying some fixed set of rewriting rules. In its simplest form, a rewriting rule is given by a pair of strings, representing a source string (the “original”) and its substitute. By a rewriting dictionary, we mean a finite list of such pairs; dictionary-based text rewriting means to replace in an input text occurrences of originals by their substitutes. We present an efficient method for constructing, given a rewriting dictionaryD, a subsequential transducer that accepts any texttas input and outputs the intended rewriting result under the so-called “leftmost-longest match” replacement with skips,t'. The time needed to compute the transducer is linear in the size of the input dictionary. Given the transducer, any texttof length |t| is rewritten in a deterministic manner in timeO(|t|+|t'|), wheret' denotes the resulting output text. Hence the resulting rewriting mechanism is very efficient. As a second advantage, using standard tools, the transducer can be directly composed with other transducers to efficiently solve more complex rewriting tasks in a single processing step.

Since the Web by far represents the largest public repository of natural language texts, recent experiments, methods, and tools in the area of corpus linguistics often use the Web as a corpus. For applications where high accuracy is crucial, the problem has to be faced that a non-negligible number of orthographic and grammatical errors occur in Web documents. In this article we investigate the distribution of orthographic errors of various types in Web pages. As a by-product, methods are developed for efficiently detecting erroneous pages and for marking orthographic errors in acceptable Web documents, reducing thus the number of errors in corpora and linguistic knowledge bases automatically retrieved from the Web.

We describe a new corpus collected for comparative evaluation of OCR-software and postcorrection techniques. The corpus is freely available for academic groups and use. The major part of the corpus (2306 files) consists of Bulgarian documents. Many of these documents come with Cyrillic and Latin symbols. A smaller corpus with German documents has been added. All original documents represent real-life paper documents collected from enterprises and organizations. Most genres of written language and various document types are covered. The corpus contains the corresponding image files, rich meta-data textual files obtained via OCR recognition, ground truth data for hundreds of example pages, and alignment software for experiments.

Character sets for Eastern European languages typically contain symbols that are optically almost or fully identical to Latin letters. When scanning documents with mixed Cyrillic-Latin or Greek-Latin alphabets, even high-quality OCR-software is often not able to correctly separate between Cyrillic (Greek) and Latin symbols. This effect leads to an error rate that is far beyond the usual error rates observed when recognizing single-alphabet documents. In this paper we first survey similarities between Latin and Cyrillic (Greek) letters and words for distinct languages and fonts. After briefly introducing a new and public corpus collected by our groups for evaluating OCR-technology over mixed-alphabet documents, we describe how to adapt general algorithms and tools for postcorrection of OCR results to the new context of mixed-alphabet recognition. Experimental results on Bulgarian documents from the corpus and from other sources demonstrate that a drastic reduction of error rates can be achieved.

The need to correct garbled strings arises in many areas of natural language processing. If a dictionary is available that covers all possible input tokens, a natural set of candidates for correcting an erroneous input P is the set of all words in the dictionary for which the Levenshtein distance to Pdoes not exceed a given (small) bound k. In this article we describe methods for efficiently selecting such candidate sets. After introducing as a starting point a basic correction method based on the concept of a “universal Levenshtein automaton,” we show how two filtering methods known from the field of approximate text search can be used to improve the basic procedure in a significant way. The first method, which uses standard dictionaries plus dictionaries with reversed words, leads to very short correction times for most classes of input strings. Our evaluation results demonstrate that correction times for fixed-distance bounds depend on the expected number of correction candidates, which decreases for longer input words. Similarly the choice of an optimal filtering method depends on the length of the input words.

Postcorrection of OCR-results for text documents is usually based on electronic dictionaries. When scanning texts from a specific thematic area, conventional dictionaries often miss a considerable number of tokens. Furthermore, if word frequencies are stored with the entries, these frequencies will not properly reflect the frequencies found in the given thematic area. Correction adequacy suffers from these two shortcomings. We report on a series of experiments where we compare (1) the use of fixed, static largescale dictionaries (including proper names and abbreviations) with (2) the use of dynamic dictionaries retrieved via an automated analysis of the vocabulary of web pages from a given domain, and (3) the use of mixed dictionaries. Our experiments, which address English and German document collections from a variety of fields, show that dynamic dictionaries of the above mentioned form can improve the coverage for the given thematic area in a significant way and help to improve the quality of lexical postcorrection methods. 1.

In this paper, we describe a new method for constructing minimal, deterministic, acyclic finite-state automata from a set of strings. Traditional methods consist of two phases: the first to construct a trie, the second one to minimize it. Our approach is to construct a minimal automaton in a single phase by adding new strings one by one and minimizing the resulting automaton on-the-fly. We present a general algorithm as well as a specialization that relies upon the lexicographical ordering of the input strings. Our method is fast and significantly lowers memory requirements in comparison to other methods.