Some frequently asked questions about using rhythm measures.
I want to use PVI to compare the rhythm of two varieties - what do I need to know?
First of all, there is large variation between different people and different texts. If you only have one speaker per language/variety, your results willonly tell you about the difference between those particular people. If you only use one text in your experiment, you will learn only about the rhythm of that particular text.To obtain results about a dialect or language, you need to have data from several speakers.
Even with a few speakers and a few texts, you will most likely be only able to speak about the general trends: expect large overlap in values between different languages. Our results indicate that there are (for instance) some English speakers who speak with French-like rhythms, some Russian Speakers who speak with a typical Mandarin rhythm, et cetera. Thus, one cannot reliably pick a measurement and deduce that "this came from a Frenchspeaker". While each language has a somewhat different "typical" rhythm, those differences only emerge when you look at a group of measurements.
How many rhythm measures are there?
There are a dozen published measures. You can find the full list with references in our paper.
Many of these measures are highly correlated with each other. Our research has shown that if you know the values offour or five of them, you can accurately predict the rest.
So which is the best one?
Our results show that there is no single "best" measure: the choice and the number of the best measures depends on the languages/varieties in your data. You may need to try several measures or combinations before you achieve the best separation. VarcoΔV, VnPVI and %V have been found most effective across several studies, but they are by no means universally the best.
Remember: if you you only find differences in non-normalized measures, you may be dealing with differences in speech tempo.
What shall I do with pauses/hesitations?
There are differentways of dealing with pauses: you can compute your RMs over each inter-pause stretch and then take the average, you can ignore pauses altogether or you may omit the final syllables. In all our corpora different ways of computation generally did not affect the results. However, this would depend on the data.
I get different values from the ones reported in N. What did I do wrong?
Nothing (Unless you analyzed the same recordings with the same labels). The difference in RMs between different speakers of the same language can be as large as the difference between speakers of different languages.
How can I reference this FAQ in my work?
We suggest you refer to our paper which presents these results in more detail:
Loukina, A., Kochanski, G., Rosner, B., Shih, C., Keane, E. (Submitted) Rhythm measures and dimensions of durational variation in speech. Preprint
You will also find references to other papers.