Problem
Kurdish Sorani has far fewer labeled speech resources and mature pretrained tools than high-resource languages, making accurate transcription substantially harder.
06 / CASE STUDY
Low-resource language · Speech recognition
Kurdish Sorani STT
A speech-to-text system focused on Kurdish Sorani, tackling the data and modeling constraints of an underrepresented language.
RoleMachine learning engineer
StatusApplied speech AI project
TimelineSpeech research and implementation
01 / CONTEXTThe problem
The problem
behind the product.
Approach
The work centered on preparing the audio pipeline, adapting recognition models and building an interface that made the resulting transcription capability usable.
02 / SYSTEMArchitecture,
Architecture,
not magic.
01Audio capture→
02Preprocessing→
03Acoustic model→
04Language decoding→
05Transcript output
Constraints
- Limited labeled data
- Dialect and speaker variation
- Audio quality
- Model adaptation
My ownership
- Prepared speech data and preprocessing
- Adapted the recognition workflow
- Evaluated transcription behavior
- Integrated the model into a usable application
03 / EVIDENCEWhat can be
What can be
defended.
End-to-end speech pipeline
Work on an underrepresented language
Model-to-interface integration
Foundation for accessible transcription
What I would improve now.
The strongest next step would be a transparent benchmark set covering speakers, environments and dialect variation before making accuracy claims.
Speech recognitionNLPDeep learningPythonAudio processing