Concatenative Speech Synthesis For Amharic Using Unit Selection Method

Speech synthesis takes text as input and generates acoustic signal as output. In the process, therninput text is preprocessed to tokenize it into words or other meaningful tokens and to transliteraternnumbers, abbreviations and acronyms. Text-analysis follows text preprocessing to identifyrngrammatical structures and context. Once the text analysis phase is completed the next step is tornconvert graphical representation of sounds to their phonetic representation. A phoneme usuallyrnhas multiple phones that are used in different contexts.rnAmharic language’s orthography is phonemical in the sense that a grapheme represents exactlyrnone phoneme. However, this statement is true as long as epenthesis and geminations are notrnconsidered. The language’s orthography does not also show suprasegmental information that isrnrequired to properly model speaking styles. Even though converting grapheme to phoneme isrneasy in Amharic, converting phoneme to phone is very difficult because of the two necessary andrnyet orthographically unrepresented components of the language – epenthesis and gemination.rnModeling prosodic features of various speaking styles is also the other challenging task inrndeveloping Amharic TTS. This is challenging because, in one hand, the task of modeling humanrnspeech is very challenging in itself and in the other hand, research works done for Amharicrnlanguage are relatively few.rnThis project work has tried to address epenthesis and gemination, which are phonologically veryrnimportant features of the language, by studying and implementing techniques found in variousrnliteratures. Making use of orthographic property of verbs in their perfect form, this workrnintroduces rules that can be used to locate phones that need to be stressed. The grapheme tornphoneme conversion algorithm also addresses epenthesis. Prosodic differences of declarative andrninterrogative utterances are represented by making use of unique sentence-final phones recordedrnand segmented for this purpose. Transliteration of numerals and abbreviations is also addressedrnin the text preprocessing phase of the system.rnThe results found after being evaluated by ten fluent speakers of the language are encouraging.

Get Full Work

Report copyright infringement or plagiarism