Swiss-German-Speech-Data---Meeting-Notes.md

@@ -2,20 +2,32 @@
 
 ### Participants
 
-Yixuan Xu, Daniel Perruchoud, Michael Graber
+Yixuan Xu, Roberto Salomone, Markus Gafner, Daniel Perruchoud, Michael Graber
 
 ### Previous Action Items
 
-- [ ] @Roberto: Confirm huggingface dump for 100 Sekunden Wissen, Licence for hugging face cc by 4.0 nc (?)
+- [ ] @Roberto: Confirm huggingface dump for Espresso, Licence for hugging face cc by 4.0 nc (?)
 - [x] @Michael: Move meeting to new slot, as requested by Roberto
 
 ### Discussion Items
 
+* Roberto expects licence draft by SRG legal team next week
+* Yixuan reconfirms that only commercially usable data can go into Apertus
 * Planning user evaluation of transcripts
+  * Markus, head of indexing @SRF 
+  * 5 people \~ 30 Minutes should be feasible
+  * sending link to Markus as soon as available
+* Preprocessing
+  * chunking for longer audio files
+  * 20 mins episode -\> 48k tokens, maybe semantically segmentation sensible -\> Yixuan will discuss with Imanol
+  * 16k tokens can be used for long context training
+  * ablation intended to be done
 
 ### New Action Items
 
-* [ ] \- 
+* [ ] Yixuan: discuss with Imanol how long srg broadcasts should be chunked, if at all
+* [ ] Roberto will provide Regionaljournal broadcasts
+* [ ] FHNW will prepare spectrograms of Espresso and Regionaljournal and use whisper for music detection if feasible
 
 ## Meeting, January 8, 2026