Developing an Embosi (Bantu C25) Speech Variant Dictionary to Model Vowel Elision and Morpheme Deletion

Abstract

This paper investigates vowel elision and morpheme deletion in Embosi (Bantu C25), an under-resourced language spoken in the Republic of Congo. We propose that the observed morpheme deletion is morphological, and that vowel elision is phonological. The study focuses on vowel elision that occurs across word boundaries between the contact of long/short vowels (i.e. CV[long] # V[short].CV), and between the contact of short/short vowels (CV[short] # V[short].CV). Several different categories of morphemes are explored: (i) prepositions (ya, mo), (ii) class-noun nominal prefixes (ba, etc.), (iii) singular subject pronouns (ngá, nO, wa). For example, the preposition, ya, regularly deletes allowing for vowel elision if vowel contact occurs between the head of the noun phrase and the previous word. Phonetically motivated speech variants are proposed in the lexicon used for forced alignment (segmentation) enabling these phenomena to be quantified in the corpus so as to develop a dictionary containing relevant phonetic variants.

Building an ASR System for Mboshi Using A Cross-Language Definition of Acoustic Units Approach

“…The Mboshi (Bantu language spoken in Congo-Brazzaville) corpus [9] consists of 5k speech utterances (approximately 4 hours of speech) in Mboshi aligned to French text. The data set also contains linguists' transcriptions in Mboshi in the form of a non-standard graphemic form close to the language phonology [10] [11] [12]. In our experiments, we used the 2087 training utterances for retraining the Baseline model and the 230 validation set utterances for which mono-phone forced alignments were available from [11].…”

“…The number of different Dutch phones in CGN is 43, while Mboshi has 68 different phones (see for more detail on the Mboshi phone inventory [12]). The output layer of the baseline model trained on Dutch thus needs to be adapted in several ways.…”

A Very Low Resource Language Speech Corpus for Computational Language Documentation Experiments

“…The word-positionindependent GMM-HMM monophone models are trained using the STK tools at LIMSI (Lamel and Gauvain, 2015). A set of 68 phonemes are used to represent the pronunciation dictionary, with multiple symbols for each vowel representing different tones (Cooper-Leavitt et al, 2017b;Bedrosian, 1996) and a symbol for silence. The acoustic model is estimated iteratively, with 5 rounds of segmentation and parameter estimation, and the model resulting from the last iteration was used to resegment the data.…”