Text to speech

PHONEMES is a program which converts the orthographic representation (the "spelling") of the input to ModelTalker into its phonemic representation (its "dictionary pronunciation"). For instance, PHONEMES converts the word "name" into NN EI MM, symbols which represent the three phonemes in the word "name". PHONEMES does not use an actual computer dictionary in order to perform its task. Instead, it goes through a complex series of context-sensitve rules (of the sort routinely used by generative phonologists) which instruct the program how to convert the letter sequence into the correct sequence of phonemes. Since the spelling of English words is quite complex and irregular, PHONEMES requires a few thousand rules to accomplish this task successfully. These rules are divided into eleven sets or "levels", which PHONEMES executes in a prescribed sequence. The first 7 levels perform a thorough morphological analysis of the input word. Of these, the first three levels detect the presence of suffixes; the next three determine whether the word is a compound; and the last level looks for common prefixes. The next two levels perform the actual letter-to-phoneme conversion, resulting in a preliminary phonemic representation. The last two levels convert this initial representation into a more detailed phonetic representation, dividing the phonemes into syllables, calculating the stress pattern of the word, applying allophonic rules (determining, for instance, which occurrences of /t/, /p/, and /k/ should be aspirated), and reducing unstressed vowels to schwas. Each level has an "exceptions dictionary" which allows exceptions to the general patterns to be fixed. In the past year, we have been working intensively to improve the rule sets so that ModelTalker pronounces more words correctly. This is being accomplished by running PHONEMES on the contents of a massive online dictionary (with more than 100,000 words), and comparing the output of PHONEMES with the correct pronunciations of these words (as recorded in the dictionary). We examine the output for patterns in the errors that PHONEMES made, and these patterns form the basis of new rules added to PHONEMES's system. In the last year, we have fixed the transcription by PHONEMES of almost 10,000 words with this process, and improved the transcription of thousands more. You might be wondering: instead of using such a complex program, why not simply look up each word in a dictionary in the computer's memory? First of all, our program takes up less computer memory than a dictionary. The program needs to store only a few thousand rules, compared to the hundreds of thousands of words recorded in a dictionary. A second reason is that the English language is constantly growing. New words are added to the language every year; English has the capacity for an infinite number of words. No dictionary could possibly hold all these words, and a program that depends on a computerized dictionary will fail badly when given a new word. In contrast, our program has the capacity to take a reasonable guess at the pronunciation of a word it has never encountered before.

Projects | Publications | Related Links | Staff |
Facilities | Events | Positions