Why do you think that a simple z-normalization does not do the job? That's what I mostly see in the literature.

Thanks for your thought, Amanda! Yes, the z-normalisation is often performed and a more sophisticated variant of 3) above (in that it takes the spread of phone durations into account). I will count your answer as a vote for 3). 

As for your question: I do not think that a simple z-normalization does not do the job. All of the normalization procedures do the job in some way. However, the z-normalization is not perfect either. For instance, it also leaves many degrees of freedom (which sounds should be used for z-normalisation: all phones in the utterance, the phones up to the target sound, the phones in the syllable or the word itself?) and is not immune to the theoretical problem I stated. 

In a production experiment one could check different normalization variables for their effect on the dependent variable (maybe also in interaction with the independent variable) - model comparisons via explained variance. This would let you identify which if any of the speech rate normalization variables does in fact affect your experimental outcome.

For different speech materials (and languages ?) different normalizations might be most appropriate. If you really want to know which normalization is most predictive of your data I think there is no way of getting around modelling the data with different normalization approaches and compare the outcomes.

I usually chose the old-fashioned syllables/second normalization. So you can count my answer as a vote for 1.

Hej Bettina,

like Xaver above I would also say that there is not general answer to your question. Different material and research questions require different normalization solutions.  For instance, if you elicit target units in a constant sentence frame, then the sentence frame duration is probably best reference for normalization. But, if you refer to spontaneous speech, small and more adjacent units could be more suitable. I guess there are two rules of thumb:

(1) Like for F0-peak alignment, neighboring units are better timing references than distant references.

(2) The target is probably represents an independent variable and is hence subject of variation. So, good/proper normalization units are those that show no or less variation than the target and/or that are common to all targets (like in the sentence frame example above)

Is this answer too unspecific?

Have a nice weekend, Oliver.

Hi Bettina,

I think there is no simple solution. It may depend on what you study. If you are interested in the tense-lax vowel contrast in German, e.g., and you have minimal pairs, I would believe taking the syllable or morpheme or word is fine - and it is very likely that the lengthened vowel coincides with a shorter consonant production and vice versa. If you want to investigate prosodic boundaries and their lengthening you may normalize with respect to the syllable/word which is preceding and following the boundary.

Recording speech rate variations, I guess, won't do the job, because affects phonemes differently (as e.g. shown in the PhD by Hartmut Pfitzinger - you can find this on the webpage of the Phonetics Institute in Munich). Vowels maybe reduced when one talks faster, but voiceless fricatives not at all.

If you are more interested in time series, dynamic time warping may do a good job.

Theoretically I know almost nothing about this - except that most phnological models do not consider time as relevant (which I think is wrong). People around Kelso often talked about the intrinsic timing - i.e. timing is a byproduct of amplitude and velocity of the articulators moving from one target to another. Not aware on any advanced theory on timing - and it may also be correlated to dozens of factors (if you look into the work on rhythm). I see your point - it is crucial how one normalizes. In the worst case - show different, even counteracting normlizations - they may already be telling what is going on between the segmental, prosodic and word level. And what people perceive is then another thing and may also work hand in hand with pitch or intensity.

Best Susanne

One warning - when doing multiple normalizations - don't run correlations on these.

http://scitation.aip.org/content/asa/journal/jasa/80/6/10.1121/1.394302

(Note on the use of correlations in speech timing. by Christian Benoit)

I think it makes sense to compare different statistical models as Xaver proposes, but be aware of potential partial correations. I guess you are anyway.

Ciao susanne

Hi Bettina,

if you have enough data from your subjects, z-scoring, but separately for each phoneme class, may be an option. I've used phoneme-specific z-scores for recognizing prosodic boundaries and pitch accented syllables. This takes into account Hartmut's finding, mentioned by Susanne above, that different phonemes are affected differently. Nick Campbell, by the way, came up with the term elasticity -- phonemes differ in their elasticity, and he introduced the phoneme-specific z-scores to model phoneme durations in synthesis in a paper in 1992.

Antje

Hi Bettina,

I basically agree with the previous answers and I also normalize my data but I wonder if normalization is always needed, i.e. if we don't lose an important piece of information by using normalization. Variation might have different sources. For instance, lengthening might be due to lexical/sentence stress, boundary lengthening, and other parameters we are not aware of.  If we normalize all this we are at risk of losing some info. In other words, is normalization always appropriate? Any thoughts on this?

Best,

Marzena

Thank you all for your thoughts and suggestions on duration normalization. You are right that the kind of normalization depends a lot on the research question and the amount of data available for normalization. Currently, I am interested in how a phonological length contrast affects the duration of adjacent sounds and possibly, like Marzena said, this is a question for which the raw data may be more useful than any kind of normalization. I wish you all a nice weekend! Bettina