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

@@ -13,7 +13,19 @@ Yixuan Xu, Daniel Perruchoud, Michael Graber
 
 * Yixuan shares https://swiss-ai.github.io/benchmark-audio-tokenizer/
   * Checkpoints of ablations
-* Transcription Evaluation for SRF broadcast transcripts
+  * SPC_R seems to only contain 482 h after VAD
+    * TODO at FHNW -\> clarify SPC_R size
+  * SPC_R has a great influence on performance for cross-modality alignment, but it's too little paired data
+  * stage 2 training: 30% transcription, 70% continuation
+  * possible solutions
+    * transcripts of Gemeinderat Zürich? -\> not available
+    * Synthetic Swiss German data?
+  * Daniel: Idiotikon data might be an option 
+  * Permissive licence
+* SRF -\> Ablation study could be done (\~ 2000 h)
+
+
+* FHNW: Transcription Evaluation for SRF broadcast transcripts
   * FHNW models perform substantially better than current SRF solution
   * ![TranscriptionRatingDistributionSRFvsFHNW.png](uploads/f4f0cc084aa9daae243cf90cb83e3485/TranscriptionRatingDistributionSRFvsFHNW.png){width="504" height="360"}
   * ![TranscriptionModelRatings_FHNW-1vsSRF.png](uploads/1a60f756ecd7ce368cebef5a11aabeb0/TranscriptionModelRatings_FHNW-1vsSRF.png){width="461" height="346"}
@@ -40,6 +52,8 @@ Yixuan Xu, Daniel Perruchoud, Michael Graber
 
 ### New Action Items
 
+* Clarify size of SPC_R dataset
+
 ## Meeting, March 4, 2026
 
 ### Participants