Govtûgô — Sorani Kurdish Speech Recognition

I fine-tuned and released Govtûgô-ASR-Sorani at RegaLabs to transcribe Central Kurdish (Sorani / سۆرانی) speech. My work builds on the existing rzgar/qwen3-asr-sorani-kurdish-ckb-v1 checkpoint, derived from Qwen/Qwen3-ASR-1.7B.

View the model and model card on Hugging Face.

My contribution

• Additional Sorani adaptation with LoRA: rank 64, alpha 128, and 196 adapted modules.
• Training run with 75,927 training utterances, 256 validation utterances, and two epochs, as recorded in the release metadata.
• Merged adapter into standalone SafeTensors weights with the processor, tokenizer, and inference configuration.
• Published the model card, WER/CER benchmarks, Apache 2.0 license, and upstream attribution.

Read and studio speech benchmarks

Evaluated on October 6, 2026, with 50 clips per dataset at 16 kHz, greedy decoding, and max_new_tokens=256. Results use corpus-level word error rate (WER) and character error rate (CER); lower is better.

Reference and hypothesis text were normalized with NFC, lowercase conversion, removal of diacritics, tatweel, bidi marks, and punctuation, plus unified Arabic/Kurdish letter forms and digits. The language Kurdish<asr_text> prefix was stripped from model output.

Some read-speech sets may overlap the training data; FLEURS is the most reliable read-speech figure. The model card contains the evaluation notes.

Uses and limitations

Designed for Sorani dictation, media captions, audio indexing, and meeting transcription. Heavy overlapping speech needs a separate speaker diarization pipeline. The model is intended for Central Kurdish; these benchmarks do not establish accuracy for other languages or dialects.

Release and attribution

Released under Apache 2.0, with RegaLabs contribution notices and attribution to the upstream Sorani checkpoint and Qwen model. Hugging Face hosts the weights and supporting files.