Testing on High German and dialects (Regionalsprache.de REDE)

With a fully calibrated set of binary classifiers the VokalJäger is equipped to measure floating phonetic features values ζ in speech signals.

Proof of concept: openness of back vowels

Below chart shows the result of an experiment: Here floating phonetic feature values ζ are measured for the openness of back vowels and the roundness of front vowels within the data on which the VokalJäger was calibrated, the Kiel PHONDAT Corpus – here: panel “HG1” [Kohler 1994]. The results are as expected: the ζ values show exactly the values – with some statistical variations – one would expect from the underlying vowels respectively their elementary phonetic feature values.

Interesting is the reaction of the VokalJäger to vowels, on which the algorithm was not calibrated. E.g. the sound [a] was omitted from training, but shows exactly the same behavior as sound [a:], on which the VokalJäger was trained – as expected from phonetic experience.

Another data set was fed into the VokalJäger for confirmation (here: panel “DFRA”). Those are dialect recordings from the Regionalsprache.de REDE project, which had been phonetically transcribed [Schmidt/Herrgen/Kehrein 2008 f.]. Hence two challenges could be tested: Firstly, can the VokalJäger, which was calibrated on High German, deal with dialect signals; secondly, does it produce the ζ values one would expect from the transcriptions? With the exception of the U-sounds, which may have to do with their representation in the Kiel Corpus, the results are as expected.

Figure. Values ζ for the floating phonetic feature openness of back vowels on the High German Kiel Corpus (left) and on Frankfurterisch recordings from the REDE material (right).  ζ=4 represents open, ζ=1 closed [Keil 2017, figure 85, p. 228; colored version].

 

Proof of concept: Roundness of front vowels

A similar experiment was conducted utilizing the floating phonetic feature of the roundness of front vowels. The results are as well encouraging: The ζ values end on the values, which are expected:

Figure. Values ζ for the floating phonetic feature roundness of front vowels on the High German Kiel Corpus (left) and on Frankfurterisch recordings from the REDE material (right).  ζ = 1 represents not-round, ζ = 2 round [Keil 2017, figure 85, p. 228; colored version].

TheVokalJäger was calibrated in this setting on the long front vowels – as such the results are most accurate for long vowels. The short front vowels range in with some significant deviations from their long counterparts but still lie within the ζ bands one would deem round ( ζ > 1.5) respectively not-round (ζ < 1.5).

Consequently, one may conclude, the VokalJäger indeed was decently calibrated on High German.