|Title:||Sintéiseoir 1.0: a multidialectical TTS application for Irish|
|Author:||Mícheál Mac Lochlainn|
|Institution:||National University of Ireland, Galway|
|Linguistic Field:||Applied Linguistics; Computational Linguistics|
|Abstract:||''This paper details the development of a multidialectical text-to-speech (TTS) application, Sintéiseoir, for the Irish language. This work is being carried out in the context of Irish as a lesser-used language, where learners and other L2 speakers have limited direct exposure to L1 speakers and speech communities, and where native sound systems and vocabularies can be seen to be receding even among L1 speakers – particularly the young.
Sintéiseoir essentially implements the diphone concatenation model, albeit augmented to include phones, half-phones and, potentially, other phonic units. It is based on a platform-independent framework comprising a user interface, a set of dialect-specific tokenisation engines, a concatenation engine and a playback device.
The tokenisation strategy is entirely rule-based and does not refer to dictionary look-ups. Provision has been made for prosodic processing in the framework but has not yet been implemented. Concatenation units are stored in the form of WAV files on the local file system.
Sintéiseoir’s user interface (UI) provides a text field that allows the user to submit a grapheme string for synthesis and a prompt to select a dialect. It also filters input to reject graphotactically invalid strings, restrict input to alphabetic and certain punctuation marks found in Irish orthography, and ensure that a dialect has, indeed, been selected.
The UI forwards the filtered grapheme string to the appropriate tokenisation engine. This searches for specified substrings and maps them to corresponding tokens that themselves correspond to concatenation units.
The resultant token string is then forwarded to the concatenation engine, which retrieves the relevant concatenation units, extracts their audio data and combines them in a new unit. This is then forwarded to the playback device.
The terms of reference for the initial development of Sintéiseoir specified that it should be capable of uttering, individually, the 99 most common Irish lemmata in the dialects of An Spidéal, Músgraí Uí Fhloínn and Gort a’ Choirce, which are internally consistent dialects within the Connacht, Munster and Ulster regions, respectively, of the dialect continuum. Audio assets to satisfy this requirement have already been prepared, and have been found to produce reasonably accurate output. The tokenisation engine is, however, capable of processing a wider range of input strings and when required concatenation units are found to be unavailable, returns a report via the user interface.
Add a new paper
Return to Academic Papers main page
Return to Directory of Linguists main page