Kaldi NL Versions Save

• Docker: updated base image from debian 11.6 to 11.8

Bugfix release:

• Fix: produce XML files again
• Docker: updated base image from debian 11.2 to 11.6
• Docker: added an additional Dockerfile to compile SRILM into the container
• reference entrypoint with absolute path for apptainer/singularity compatibility
• README: removed LaMachine reference

Bugfix release:

• Support for mp3 and more robust XML conversion also in case of non-wav input (e.g. mp3)
• better error propagation when speaker diarisation fails
• utwente models are also required for radboud models
• Ensure there are no trailing slashes for the output dir

• Re-containerized kaldi and Kaldi_NL (two Dockerfiles), no longer using LaMachine because that may eventually be deprecated.
  • This results in leaner containers
  • I decided to offer a kaldi image as part of this repository rather than relying on the upstream kaldi container because 1) their image is out of date (uses debian 9 whilst debian 11 is current stable), 2) we can do some more optimizations for smaller image size
  • Container builders can decide whether to install models inside the container, or on an externally mounted volume (automatically populated upon first run). This results in images of 15.5GB (with models) respectively 3.97GB (without models), the former used to be just over 32GB as optimization was not a priority. The kaldi image itself (without Kaldi_NL) is 3.64GB.
  • Pre-built images are provided via my Docker Hub (proycon/kaldi, proycon/kaldi_nl), probably to be moved to a CLARIAH registry later.
  • Note: the asr_nl webservice image is not yet completed in the new style (still uses LaMachine) and will follow later.
• Made various shell scripts POSIX-compliant and more robust
• Models are now individually installable and the default option is to provide NO models at all
• Various models shift a lot of unnecessary cruft that is now removed
• Documentation and contributor guidelines have been adapted to the new situation

• oral history has been renamed to asr_nl (since v0.5.0)

• Moved Kaldi_NL models from Twente to Nijmegen, because server in twente was down.

This release implements the following:

• Greatly improved error handling and logging; hopefully making it easier to pinpoint problems in case things do go wrong. #17
• Reduced code duplication in the oral history contributions by merging the three decode scripts into one
  • The main scripts now pass fundamental shellcheck tests
• Two functions have migrated from oral history webservice into kaldi_NL itself:
  • Added ctm2xml conversion
  • Process and output speaker diarisation information

Release of the current status-quo

This release uses a modelpack that primarily supports NNet1 Bottleneck models and has a single universal decode.sh