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

-# Meeting, November 13, 2025
+## Meeting, November 13, 2025
 
-## Participants
+### Participants
+Yixuan Xu, Seven Najem-Meyer, Imanol Schlag, Roberto Salomone, Vincent Demotz, Daniel Perruchoud, Michael Graber
 
-## Discussion Items 
+### Discussion Items 
 
-- Introductory round
+#### Introductory round
 
-### Data Requirements
+#### Data Requirements
 - Speech Freqs 16 - 24 kHz, Input for Tokenizer
-- Ideally dataset would be paired, ideal with speaker annotations
-- ~ 20k hours audio data, not only Swiss German
-- Languages, all european languages
+- Ideally dataset would be paired, even more ideal including speaker annotations and dialect labels
+- Goal: ~ 20k hours audio data, not only Swiss German
+- Audio of all European languages will be included
 - Ideally openly available, like hugging face
-- Model does not memorize
-
-Roberto
-- SRG could provide own production data, not all data
-
-Imanol
 - Data should be reproducible / versioned
+- Data needs to be available online at model release -> reproducibility
+- SRF quality should be good enough
+- All Swiss Languages from SRG are welcome
+
+Re licencing (stressed by Immanol): 
+- The model does not memorize (due to Goldfish loss)
+- Hence the model is not a derivative of the data
+- Due to lack of memorization, license constraints of data does not need to apply for the models
 - Model is released under Apache 2
-- Model is not a derivative of the data
-- Open for other data sources
-- data needs to be available online at model release
 
-### Processing steps (Yixuan)
+#### Possible Sources Roberto
+- Roberto: SRG could provide own production data, not all data
+- Apertus: Open for other data sources 
+- Synthesized data could be helpful as well
+
+#### Processing steps
 - speaker tags
 - transcriptions
 - de-duplication
 - detoxification
+- Tokenization is done by Apertus Team
 
-Model generation will be restricted to text, no audio, no images
-
-Synthesized data could be helpful
-
-Tokenization @ Apertus
-
-Infrastructure
-- Clariden
-- Swiss AI GitHub
+#### Infrastructure
+- Clariden can be used
+- Ultimately, codes should be available through Swiss AI GitHub
 
-Timeline / Roadmap
+#### Timeline / Roadmap
 - Tokenizer identification up next
-- First experiments 100B Tokens
-- The earlier the better
-
-- Soon 1000 h paired would be helpful, and worth it
-- Running audio experiments woul
-
-
-Apache license 2.0 was shared with SRG.
-
-Communication in slack, meetings on demand.
-
-From SRG all Swiss Languages are helpful
-Dialect labels and transcriptions would be 
+- First experiments 100B Tokens will follow
+- It would be helpful if ~ 1000 h paired Swiss German data would be available
 
+#### Varia
+- Model generation will be restricted to text, no audio, no images
+- Apache license 2.0 was shared with SRG.
+- Channel on Swiss-AI slack will be created
+- To start, meetings will be held weekly, same timeslot
 
-## Action Items
+### Action Items
 
+- [ ] Complete data source list (@michael.graber, @daniel.perruchoud )
 - [ ] Define roadmap (@michael.graber, @daniel.perruchoud )
+- [ ] Identify list of downloadable casts complying with Apertus co (@Roberto & @Vincent / SRF)
+- [ ] Setup meeting series in same slot (@michael.graber)
\ No newline at end of file