Minutes of the meetings and other general matters
Participants: Timo-Pekka, Jaana, Tuomas, Alex, Shriram, Eemi
We should send reminders to participants by text message and e-mail. Jaana will take care of this.
Large number of native speakers is not needed.
Adrian Benfield would listen and annotate the recordings in addition to Jaana.
For the annotation, all the instances of a word will be listened in a row.
If multiple phonemes in a word have often errors, those should be annotated. Initially let us aim at analyzing errors in one phone. If the other errors cause problems, we can ignore words where multiple phonemes are erronous.
Annotation will be done independently by Jaana and Adrian.
Jaana will first annotate 5 words which have promiment, easily recognizable errors. Those should include different target phonemes.
For each instance of a word, the annotation should include the following information:
-is the sample contains a gross error (yes/no)
-is the target phoneme correct (yes/no). if not correct, a letter describing the error
-is a non-target phoneme incorrect (yes/no). if yes, a letter describing the non-target phoneme and a letter describing the error.
TUT should start doing segmentation of words using the material we have now, and continue as we receive new material. We will use a semi-automatic approach. All the instances of a word are in a folder which name indicates the word. Files of segmented words can use a running number from 1 to N. TUT has information that links the persons and these numbers.
Shriram will do forced alignment of the material, and check if temporally segmenting and listening to them could be used to easily verify the correctness of segmentation.
Target to have the segmented files ready on 26th June. TUT will prepare some material for Jaana to be familiarized with the task by 13th June.
Shriram will setup a password-protected web page for the project.
Participants: Timo-Pekka, Jaana, Alex, Shriram, Eemi
Jaana updated about the general conduct of the recording sessions.
Some brainstorming: -what errors need to be annotated- errors in both target and non-target phonemes? -Or classifying errors as typical or atypical?. -Should the output be something like correct, typical error or garbage? -How to prepare the dictionary? What to encode and how to encode? Also mark the garbage cluster. -Quite confusing now, will be clarified better by email after discussing with Adrian.
How to make the annotation process easier? Explore the media player options.
Dropbox will be used for sharing annotation files.
Eemi makes sure that the recordings will be ready for annotations immediately after the midsummer break.
Shriram has started setting up the tools for algorithm development.
Participants: Timo-Pekka, Adrian, Jaana, Tuomas, Alex, Shriram, Eemi
Some annotations from Adrian are still missing. He will do those in two days.
Shriram gave a status update of the ongoing research (see the attached slides).
TUT should have a complete system for detecting pronunciation errors even though individual parts are still not fully optimized. This will give us a picture of the importance of individual system components
Shriram will calculate statistics of the annotations and annotator agreements.