Changes
Page history
Update Swiss German Speech Data Meeting Notes
authored
Mar 11, 2026
by
Michael Graber
Show whitespace changes
Inline
Side-by-side
Swiss-German-Speech-Data---Meeting-Notes.md
View page @
fddcdc28
...
@@ -11,6 +11,23 @@ Yixuan Xu, Daniel Perruchoud, Michael Graber
...
@@ -11,6 +11,23 @@ Yixuan Xu, Daniel Perruchoud, Michael Graber
### Discussion Points
### Discussion Points
*
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"}
*
!
[
TranscriptionModelRatings_FHNW-2vsSRF.png
](
uploads/fad633a86e28d58200bb69dfaa4204e5/TranscriptionModelRatings_FHNW-2vsSRF.png
)
{width="461" height="346"}
*
Rating scale:
1 - UNUSABLE: Largely incomprehensible; re-transcription would be faster than correction.
2 - POOR: General content recognizable, but frequent errors; substantial post-editing required.
3 - ADEQUATE: Essential content correct, errors in difficult passages; moderate post-editing needed.
4 - GOOD: High accuracy, errors only with rare words or poor audio quality; minor corrections needed.
5 - EXCELLENT: Nearly error-free, even with technical vocabulary and difficult conditions; practically publication-ready.
*
\-\>
will use FHNW models to generate transcripts, as soon as licencing situation is clarified
*
SFT datasets in discussion for Swiss German translation (text and audio)
*
SFT datasets in discussion for Swiss German translation (text and audio)
*
https://huggingface.co/datasets/deepset/germanquad
*
https://huggingface.co/datasets/deepset/germanquad
...
@@ -18,12 +35,6 @@ Yixuan Xu, Daniel Perruchoud, Michael Graber
...
@@ -18,12 +35,6 @@ Yixuan Xu, Daniel Perruchoud, Michael Graber
https://huggingface.co/datasets/lawinstruct/lawinstruct (have to extract german (de))
https://huggingface.co/datasets/lawinstruct/lawinstruct (have to extract german (de))
*
FHNW has to identify suitable models (not trained on proprietary data) for translation
*
FHNW has to identify suitable models (not trained on proprietary data) for translation
*
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}
*
!
[
TranscriptionModelRatings_FHNW-2vsSRF.png
](
uploads/fad633a86e28d58200bb69dfaa4204e5/TranscriptionModelRatings_FHNW-2vsSRF.png
)
{width=461 height=346}
*
\-\>
will use FHNW models to generate transcripts, as soon as licencing situation is clarified
### New Action Items
### New Action Items
...
...
...
...