No abstract

Motivated by the advent of security solutions which rely on voice biometrics, we revisit by means of extensive computer-based investigations the concept of phonetical balance for Romanian utterances. We show that the standard distribution of phonems offers only a partial description of the phonetics of the language and that more detailed statistical indicators are needed. To this end, we introduce a simple indicator that measures vowel-consonant (or consonant-vowel) sequences and analyze the distribution of consonant clusters for Romanian words. Our results show that the distribution of consonant clusters is scalefree-like (akin to the distribution of words and phrases in large texts) and that large clusters of vowels or consonants are infrequent. This, in turn, indicates that utterances consisting of words which are statistically unrepresentative with respect to the previous indicators are good candidates for benchmarking the efficency of voice biometrics solutions.

Abstract. The calibration of voice biometrics solutions requires detailed analyses of spoken texts and in this context we investigate by computational means the rank-frequency distributions of Romanian words and word series to determine the most common words and word series of the language. To this end, we have constructed a corpus of approximately 2.5 million words and then determined that the rank-frequency distributions of the Romanian words, as well as series of two, and three subsequent words, obey the celebrated Zipf law.

No abstract