Central Kurdish STT model

submit/Variant_Accent_Dialect/kaka-central-kurdish/ReadMe.md

@@ -0,0 +1,38 @@
+
+# Central Kurdish STT Project
+
+## Submission for Our voices competition
+
+
+
+Kurdish language has two major dialects, Sorani (Central Kurdish) and Kurmanji (Northern Kurdish), both of which lack a STT solution primarily because they lack a voice corpus. This, however, has changed alot thanks to the Common Voice project. Now with the latest 11th release of this corpus we have access to 100+ hours for Sorani and 50+ hours for Kurmanji. This is the model I built for this competition.
+
+
+# TLDR;🤌
+
+ - Small and efficient, 140MB checkpoint, 70MB when converted to ONNX.
+ - Faster than realtime on CPU 🔥
+ - Python-ready, C# Windows port and Word addin with live transcription on its way
+ - Results on [Asosoft corpus](https://github.com/AsoSoft/AsoSoft-Speech-Corpus) (4.30hours Sorani speech, 7 males, 1 female)
+
+| Decoder | WER |
+|--|--|
+| greedy | 0.0 |
+| kenlm| 0.0 |
+
+
+# Building blocks
+
+ - [NeMo](https://github.com/NVIDIA/NeMo)
+ - [Quartznet](https://arxiv.org/abs/1910.10261) network
+ - Finetuned a [pretrained English model](https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/stt_es_quartznet15x5) by Nvidia on CV11 Central Kurdish
+ - All validated.tsv files have been used. first 80% of files used for train and the remaining 20% for test.
+ - Unfortunately CV8-11 contains non standard Kurdish text which needed to be corrected prior to training. To this end I have used [Asosoft library](https://github.com/AsoSoft/AsoSoft-Library) to normalize the sentences. Abdulhadi Riwandizi provided some tricks which seem to have corrected almost all errors. (Thanks!)
+ - Both train and test splits I used are provided.
+ - Asosoft has released [a subset of their internal dataset](https://github.com/AsoSoft/AsoSoft-Speech-Corpus) for public use. I have tested the final model on this.
+ - The python version lacks LM based decoding. However a .NET version of the software has been developed which employs a very small KenLM model leading to lower WER.
+ - Training logs are available on Wandb.
+ - Download pretrained checkpoints from here
+ - LM boosting uses a char KenLM model I trained on [Asosoft text corpus](https://github.com/AsoSoft/AsoSoft-Text-Corpus) (o=5) weighting 31MB.
+ - Training was done using a single RTX 3090 for about 12 hours. (batch size=64)
+ - Kurdish is phonetically consistent. This might have helped the network learn faster. Given the results and small dataset used, this can be a good starting point for future Kurdish STT work.