Towards a Methodology to Estimate Minimum Sample Length for Speaking Rate

Abstract

The aim of the present work is to investigate how long a speech sample should be so that the speaking rate derived from it be considered representative of the whole utterance from which the sample has been taken. Eight Brazilian Portuguese speakers read a 144-word text in three rate levels: slow, normal and fast. Speaking rate was measured cumulatively as the number of phonetic syllables per second from the first to the last syllable in the sample. Two types of rates were measured, speech rate and articulation rate. Change point analysis was used to determine the influence of rate type and level on the amount of time necessary for the cumulative estimate of speech and articulation rates to stabilize around the rate yielded by the whole utterance. Mean stabilization latencies are 9.2 seconds for speech rate and 8.7 s for articulation rate. The slow rate tends to stabilize later than fast and normal rates for both types of rate. Stabilization intervals take up a median number of 41 (speech rate) and 59 (articulation rate) syllables. Mean deviations between the global rate and the rate value at stabilization point are 7.8% (speech rate) and 4.2% (articulation rate). The stabilization times estimated here can be useful in sociophonetic research, forensic and clinical phonetics.

Minimum Sample Length for the Estimation of Long-term Speaking Rate

“…He introduces a methodology to determine this sample size in an objective way that is described in more detail in section 2.5. In this paper we compare his methodology with the one we developed independently and described in [1]. The results presented here show that applying Kendall's stabilization criterion with some modifications inspired by our own approach the estimates that we arrive at are similar to the ones we reported before and are much shorter, in the order of less than 30 seconds.…”

“…The speech material analyzed here is the same as in [1]: a 110-word long text called "A Menina do Narizinho Arrebitado" by Brazilian writer Monteiro Lobato, read by 8 Brazilian Portuguese (BP) native speakers (3 female, 5 male) with ages ranging from 18 to about 30. This particular passage was chosen because it has the advantage of being relatively short and at the same time it contains at least one occurrence of all BP phonemes.…”

“…When pause and other disfluencies are included in the duration interval used to measure speaking rate we have speech rate and when only the duration of fluent speech is considered we have articulation rate. Following our previous work [1], we investigate both speech rate and articulation rate.…”

“…In a previous work [1] we explored a methodology to determine the minimum amount of speech data necessary to yield a stable measure of speaking rate. It consisted of obtaining the time series defined by the cumulative speaking rate (speech or articulation) calculated over an entire audio sample and submitting it to a statistical technique called change point analysis, that detects the point in time where a significant change in the underlying variance of the time series takes place (see [11] for technical details).…”

“…In this paper we expand on a previous work [1] that tried to suggest some directions on how to give a principled answer to the question of determining what the minimum audio sample length for the estimation of long-term speaking rate should be so that the estimate derived from a part of the sample can be considered representative of the whole sample and maybe of other samples by the same speakers.…”