HQSYN16

Allophonic variants¶

We performed a detailed auditory analysis of the synthesized phrases. Details are attached as a separate file, allophonic variants.docx, in Czech.

The results can be briefly summarized as follows:

1) [r] vs. [r=]
A normal [r] is very often realized as a flap (i.e., not a trill) in Czech. When such an [r] is inserted into words which should have [r=], quality is impaired - the target sound is very short and the absence of trilling is disturbing.
Sometimes there is an impression of an [r+vowel] sequence, rather than [r=].

2) [l] vs. [l=]
When a syllabic [l=] is used in contexts where a [l] would be followed by a vowel (i.e., potlach [potlax]), a glottal stop is usually inserted [potl=?ax], which is extremely disturbing.
Sometimes there is an impression of an [l+vowel] sequence, rather than [l=].

3) [n] vs. [N]
When alveolar [n] is used in pre-velar contexts, it is usually accompanied with an epenthetic schwa [len@ka].
When velar [N] is used in non-velar contexts, it is often audible and disturbing.

4) [x] vs. [G]
Here the phonetic voicing of the segment often did not correspond to the definition; none of the sequences was disturbing.

5) voiced ř [\P] vs. voiceless ř [\Q]
(Partially) voiceless realizations in contexts which should have a voiced realization is usually not disturbing, as ř is frequently (partially) devoiced in ordinary speech.
However, when the voicing of ř also changes the voicing of a neighbouring obstruent, the result is disturbing and may even lead to change of meaning (dřímá --> třímá).
When a voiced [\P] is used in contexts which should be voiceless, the result is disturbing.

I have uploaded to wiki new files with synthesized utterances - now all should be correctly synthesized. Except the fixes of missing items and more variants of [ch] replacements, the remaining files contain the same data as were in the previous version (thus they does not have to be analysed again).

The second round of auditory analyses did not reveal any new information. The conclusions and recommendations from the previous phase hold:

1) the two versions of /l/ - plain [l] and syllabic [L] - should be kept separately, as the l -> L interchange in prevocalic contexts almost always leads to synthesizing a glottal stop (e.g., [vL!ak]).

2) the alveolar [n] and velar [N] should be kept separately, the N->n interchange usually yields an epenthetic schwa (e.g., [len@ka]).

3) the two versions of ř - voiced P\ and voiceless Q\ - should be kept separately

4) the allophones of /x/ - [x G h] - can be pooled together, without any marked intrusive effects.

The same as before, but now for male voice (Jan, spkr_AJ) - data put as attachment on wiki.

Auditory analysis of the male speaker yielded more favourable results than for the female. Details are attached as a separate file, allophonic variantsJan.docx, in Czech, the main findings are summarized here:

1) based on the listening sample, the two versions of /r/ - plain [r] and syllabic [r=] - could be pooled

2) the two versions of /l/ - plain [l] and syllabic [l=] - could be pooled because the l -> l= interchange in prevocalic contexts never led to synthesizing a glottal stop (it almost always did with the female speaker).

3) the alveolar [n] and velar [N] could also be pooled, the N->n interchange led to an epenthetic schwa (e.g., [len@ka]) only once, and the effect was not very disturbing.

4) the two versions of ř - voiced P\ and voiceless Q\ - should be kept separately

5) the allophones of /x/ - [x G h] - can be pooled together, without any marked intrusive effects

In conclusion, it seems advantageous to specify the pooling or keeping apart of allophonic variants for every speaker in the database separately.

The same as before, but now for male voice (Stanislav, spkr_JS) and for female voice (Kateřina, spkr_SK) - data put as attachment on wiki.

The results of the last set of analyses mostly follow the previous findings, and they are provided in the attached documents. Here the task is summarized and conclusions are drawn:

1) Regarding the syllabic [r=] and [l=], it does not seem to be possible to generalize with respect to the speaker's sex. In the female voice, Katerina, the presence of syllabic [r=] led several times to the insertion of the glottal stop, e.g. drozd as [dr=!ost]; this never happened with the male speaker, Stanislav. However, when [l=] was synthesized, a glottal stop was inserted in most of the cases in the male speaker's voice, Stanislav, and never with the female voice.
Conclusion: preferring to err on the side of caution, syllabic and non-syllabic sonorants - [l ] and [l=], as well as [r] and [r=] - should be kept separately.

2) The interchange in synthesis of the allophones of /n/ - [n] and [N] - sometimes leads to the insertion of an epenthetic schwa (e.g., [len@ka]); the occurrence, as well as the salience of the schwa differs in individual speakers. Assuming that a nasal coming originally from the velar context would hardly be synthesized into an alveolar context in natural conditions (e.g., víno as [vi:No]), the answer to the question whether to keep the two contexts separately is not quite straightforward and depends on the richness of the segment inventory.
Conclusion: the epenthetic schwa is sometimes rather intrusive, but if the "cost" of keeping [n] and [N] as separate segments were to be too high, it is possibly to pool them.

3) The results are unequivocal with regard to the Czech fricative trill /ř/ and its too allophones - the voiced [P\] and voiceless [Q\]. Interchanging one for the other typically leads to very intrusive effects.
Conclusion: the two allophones of /ř/ must be kept separately.

4) The effect of interchanging the allophones of the velar fricative /x/ seems to be relatively minor, especially when it comes to [x] and its voiced counterpart [G]. The laryngeal [h] sometimes is intrusive, especially when it also triggers incorrect voicing assimilation; under normal conditions, this would most probably not happen, though.
Conclusion: the allophones of /x/ can be pooled together.