528 14 Speech, Text and Braille Conversion Technology

TTS systems compared to recorded speech

The availability of large capacity cheap memory chips has made it possible to store whole spoken utterances, thereby eliminating the need for TTS synthesis under the following conditions:

•The speech corpus consists of a fixed set of utterances.

•The speech corpus is sufficiently small to have limited memory requirements for storage and limited time requirements for production by a human speaker.

•The corpus is fixed, with low probability that it will be necessary to change it in the future.

Recorded speech has a number of applications in daily life, such as public announcements at stations. The principle is known from voice recorders for storing short dictations or other audio information. These recorders are available as both small stand-alone devices and integrated into pocket PCs and other applications. Recorded speech can be of hi-fi quality or reduced quality due to degradation by the codec used to compress the signal. However, recorded speech always sounds more natural than synthetic speech. This can lead to problems in evaluating the quality of TTS systems if users are not aware of the difference between recorded and synthetic speech.

14.5 Braille Conversion

14.5.1 Introduction

Braille is used to represent text by means of tactile symbols. Since tactile text reading is discussed in Chapter 4, it will only be considered briefly here. The main idea was developed (from an unsuccessful system developed for the French army and called night writing) by the French blind teacher Louis Braille in about 1825 and aimed to replace the visible characters in text by a tactile array of up to six dots. Braille is still an important means for blind people to access written documents, though the proportion of blind people who read Braille is low. The elements of a printed text are called characters or signs, whereas the elements of a Braille transcript are called symbols.

Braille code is physically implemented in the form of small raised bumps on paper or other material, which are touched with the fingertips. The bumps can be produced using a stylus and a slate. The slate comprises an upper part with rectangular openings to ensure appropriate positioning of the symbols, and a lower part with small indentions for positioning the dots. Producing large texts in this way is both time-consuming and complicated, as a mirror writing approach must be used for punching the Braille symbols to produce raised bumps rather than depressions. Therefore, mechanical writers called Braille writers or Braillers are used. They are similar to a typewriter with six keys (one key for each dot of a Braille symbol) and a space bar, as illustrated in Figure 14.18. They are operated by Braille transcribers,

Figure 14.18a,b. Old and new Braille writing devices: a mechanical writer, embossing on a paper tape for stenographic purposes (Firma Karl Neubert, Leipzig, around 1950), from the historic collection of the TU Dresden (photograph by R. Dietzel); b electronic Braille writer Elotype 4E, also applicable as computer-controlled Braille embosser (by courtesy of Blista Brailletec, Marburg)

who are usually certified professionals, as they have to perform the transcription process mentally, which is not carried out by an automatic (technical) system.

Automatic systems for transcription from text into Braille and Braille into text have a number of advantages, particularly with regards to speeding up the process, but require suitable input and output devices. Other than the need for appropriate hardware, the conversion seems relatively simple, as it is performed sign by sign at the text level. As will be pointed out in the following, this is not in fact totally the case.

The original approach was based on a one-to-one coding of the characters with code combinations defined for all letters, numbers, punctuation marks, and a few special symbols, such as those indicating capitalization. This baseline system, which includes a few elementary composite signs, is called Grade 1 Braille. Producing text in Grade 1 Braille requires a lot of space. One page of a Braille book contains approximately 1000 Braille symbols, whereas one page of a printed book may contain 3500 characters. Therefore a number of rules were introduced to reduce the volume of a Braille coded text by 20–30%. (This is comparable to using shortcuts such as “asap” for “as soon as possible” or the sign “4” for the word “for” in SMS or e-mails.) This has resulted in an extensive system of so-called contractions and the rules for their application. For instance, if “&” can be used for “and”, then “h&le” stands for “handle”.

The complete system is called Literary Braille Code, Grade 2 Braille or contracted Braille. Most blind people who read Braille can also read the contracted version. Unfortunately, the contractions are language specific and therefore the code for contracted Braille must be learnt for each language separately. In German speaking countries, the terms Vollschrift (full text) and Kurzschrift (short text) are used for Grade 1 and Grade 2 Braille. Different organizations coordinate the standardization in the different countries, as shown in Table 14.7. Different approaches are required in some non-European language systems. For instance, Chinese Braille symbols represent the sounds of the spoken language rather than the characters of the written language.

530 14 Speech, Text and Braille Conversion Technology

Table 14.7. Examples for organizations and resources of the Braille system for different languages

aA general overview on the codes in various languages is presented in homepages.cwi.nl/ dik/english/codes/braille.html

There are also special forms for the following purposes, amongst others:

•Musical notation (www.brl.org/music/index.html; for details see Chapter 16).

•Mathematical notation; although there is a universal mathematical text notation, this is unfortunately not the case for Braille and different systems of mathematical Braille have developed in different countries. This has a number of disadvantages, including making both the production of mathematical Braille and communication between blind mathematicians in different countries more difficult. The US version is known as The Nemeth Braille Code.

•Scientific notation, for instance for chemistry (www.brl.org/chemistry/).

•Phonetic notation (www.clauchau.free.fr/L/phonalph.html).

•Computer Braille, which uses an 8 dot system for 256 signs, analogously to the ASCII code (see www.braille.org/papers/unive/unive.html or DIN 32 982).

With regards to text-to-Braille conversion, there are clearly additional difficulties associated with the production of contracted Braille. Some remarks on this topic can be found in the first volume of this series (Hersh and Johnson 2003, pp 262– 265). The following two subsections consider hardware and software aspects of text-to-Braille and Braille-to-text conversion. It is often useful to provide both Braille and speech output and examples of dual output systems are presented in Section 14.6.

14.5.2 Text-to-Braille Conversion

Historical development

Analogously to speech-related technologies, the development of modern reading machines has required the availability of computer technology. However, the requirement for tools for blind people to read printed texts and thus enhance their independence was recognized decades earlier. The discovery of the photoelectric effect allowed printed signs to be converted into electrical signals, which were presented in audio form in the earliest reading devices. Early devices with audio output included the following:

•The Optophone was invented in 1914 by E.E. Fournier-D’Albe, a Professor in Birmingham, UK. In the optophone the printed text characters are irradiated with light pulses and the reflected light is converted into an audible signal by means of a Selenium cell. The listener uses the different sounds to distinguish between the different text symbols. The invention was subsequently modified and applied in several different devices.

(See www.oldweb.northampton.ac.uk/aps/eng/research/optophone/ optophone2.html.)

•Another early reading machine was constructed by Rosing, a Professor in Petersburg before 1917. The text characters were scanned by an optical beam in the horizontal and vertical directions to produce a Morse-like sound.

The complexity and lack of an intuitive relationship between the text and the acoustic patterns and the lack of synthetic speech output led to the development of reading devices with a tactile output, including the following:

•W. Thorner, an eye specialist in Berlin, obtained a patent in 1916 for converting images to electro-tactile patterns. Later inventions used grids of electromagnets for the production of mechanical stimuli. This approach is still used for presenting graphical information (see Chapter 4 of this book). However, it proved to be unsuitable for reading text unless the magnetic elements were combined to give Braille symbols.

•G. Schutkowski, a teacher at a school for blind people near Berlin, was probably the first person to recognize that the most appropriate tactile output for reading machines was Braille. He invented an optical pattern matching method for reading (Schutkowski 1938). The reading mechanism was coupled with a drum which moved the corresponding Braille symbol in front of the user (Figure 14.19a). Remarkably, he also investigated the addition of speech output to his device (Schutkowski 1952) and realized that the spelling mode which he developed was only the first step towards a full text-to-speech system.

The current situation is summarized in Figure 14.20. Technical support systems for a human transcriber are presented in the upper part of the figure. The Brailler has already been mentioned. It can be replaced by a PC with driver software which changes the standard keyboard to a six-key input device. Automatic text-to-Braille conversion is illustrated in the lower part of the figure. The components of this process will now be described.

532 14 Speech, Text and Braille Conversion Technology

Figure 14.19a,b. Old and new devices for volatile Braille output: a the drum of the early prototype of a text-to-Braille machine by G. Schutkowski (contemporary press photograph from the early 1950s); b a Braille display with 40 elements (type SuperVario), mounted in front of a PC keyboard (by courtesy of BAUM Retec AG)

Input devices

The text to be converted needs to be available in machine-readable form. If this is not the case, a text file can be produced in one of the following standard ways:

•By using a PC with a conventional text processing system.

•By using a scanner to input the printed text. The scanner is connected to a standard PC which hosts the optical character recognition (OCR) software which produces the text file. Since errors cannot be completely avoided with the current state of OCR technology, the word processing software of the PC should be used to edit the resulting file and correct any errors.

The scanner is the essential input device for a reading machine. Since it performs similarly to other OCR applications, its properties will not be discussed here. They are described in more detail in Chapter 15 and the reader is referred to Bunke and Wang (1997) for additional information.

Output devices

The output device used depends on whether a durable but bulky paper version of the converted text or a quick to produce but transient (volatile) version is required.

Producing Braille symbols on paper using an automatic system requires the manual keys of the traditional Brailler to be replaced by a corresponding computer interface, called a Braille embosser. A number of different types of embossers are available and an illustration is presented in the first volume of this series (Hersh and Johnson 2003, p 265). See also Chapter 12 of this book. Some types can produce dots on both sides of the page (interpoint Braille). An electronic Brailler can be used as a computer output device also (Figure 14.18b).

534 14 Speech, Text and Braille Conversion Technology

Braille displays are used to provide the user with refreshable (volatile) Braille output. They consist of cells which each produce a touchable Braille symbol by using a piezoelectric drive to raise a number of small pins. Displays consist of lines of 20, 40, or 80 cells, as shown in Figure 14.19b. Braille displays were introduced by the company F.H. Papenmeyer in 1975. Modern Braille cells have 8 dots and can therefore be used for computer as well as literary Braille.

Conversion software

Text-to-Braille conversion software is the main component of Figure 14.20. It needs to carry out the following different groups of tasks:

•Mapping the text characters to Braille symbols. This is the central module of the conversion process. It should be noted that this mapping is not one-to-one mapping to the use of a large number of contractions and special rules, which are language-dependent as it was explained already.

•Text analysis. The use of contractions is governed by a number of specific rules. These include shortcuts for prefixes and suffixes, preference rules for ambiguous sequences and contractions which cannot be applied over syllable boundaries. Therefore, the text-to-Braille mapping module needs to “know” the morphological structure of the words and this is the task of a special module. Morphological analysis is a typical AI problem, which is apparently relatively easy to carry out for a person, but difficult for an automatic system. The accuracy of the text-to- Braille conversion depends on the morphological analysis. However, there are frequently morphological ambiguities, which can only be resolved only at the semantic level. Morphological analysis is also a component of the text analysis module of a TTS system, as shown in Figure 14.15. In both cases, certain types of errors are difficult to avoid and lead to prosodic errors in TTS and transcription errors in text-to-Braille conversion.

•Producing the correct layout from the Braille symbols. There are clearly defined layout rules for books and other documents in Braille in order to facilitate navigation and structure recognition by blind readers. The conversion software must also consider these layout rules, particularly if an embosser is used to produce a paper document. The layout rules for English Braille documents are summarized in the document “Braille Formats: Principles of Print to Braille Transcription 1997” (www.brl.org/formats/).

As shown in Figure 14.20, the conversion software works on a file-to-file base. This has two main advantages and one drawback:

•Using a text file for input is convenient for storing and re-using the data and is also required to enable removal of any errors introduced by the scanner OCR software using word processing software. This is easier than correcting the Braille output.

•The Braille file can be evaluated and corrected using special software tools. This is necessary to remove errors from the text-to-Braille conversion. These