Describe the models

README.md

@@ -0,0 +1,25 @@
+---
+license: other
+license_name: per-file
+language:
+- sr
+tags:
+- whisper.cpp
+- gguf
+---
+
+# LearnLanguage models
+
+Models the LearnLanguage app (a language-learning app) downloads to run on the phone.
+Each file is pinned by the app to a commit of this repo. Licenses differ per file:
+
+## whisper.bin — Serbian speech recognition
+
+whisper.cpp model (ggml, q5_0) of [openai/whisper-large-v3-turbo](https://huggingface.co/openai/whisper-large-v3-turbo),
+fine-tuned for Serbian with LoRA (merged): transcripts in Latin script with case and punctuation.
+Word error rate (whisper.cpp, greedy): Common Voice sr, held-out speakers — 9.4% (base model 39.8%);
+FLEURS sr_rs test — 12.1% (base model 13.3%).
+
+- License: MIT, as the base model.
+- Training data: [FLEURS](https://huggingface.co/datasets/google/fleurs) sr_rs (CC BY 4.0, Google)
+  and [Common Voice](https://commonvoice.mozilla.org) 27.0 sr (CC0, Mozilla).