Kurdish Voices

Dialect-specific automatic speech recognition for Kurdish

Datum: seit 2024
Status: aktiv

Überblick

Fine-tuning automatic speech recognition models on Kurdish dialect data to enable dialect-specific ASR solutions for low-resource spoken varieties.

Overview

Kurdish Voices is a project focused on building dialect-aware automatic speech recognition (ASR) systems for Kurdish. Existing multilingual ASR models - even those that nominally support Kurdish - are typically trained on a narrow slice of one dialect, meaning speakers of other varieties receive significantly degraded recognition quality. This project investigates targeted fine-tuning strategies that allow a base model to specialise for a specific Kurdish dialect while retaining robustness.

Motivation

Spoken language is the primary mode of communication for many Kurdish speakers, particularly in rural and diaspora communities where literacy in Kurdish may be limited. High-quality dialect-specific ASR is therefore not merely an academic goal - it is a prerequisite for applications like voice interfaces, automatic transcription of oral history recordings, and accessibility tools. The current gap in dialect-aware Kurdish ASR is significant and largely unaddressed in the research literature.

Approach

Fine-tuning pre-trained multilingual speech models (e.g. Whisper-family) on dialect-specific Kurdish audio data.
Leveraging the dialect-specific text corpora produced by the Dialectal Filtering project as a language modelling resource.
Evaluating recognition quality per dialect using appropriate word error rate (WER) benchmarks.
[TODO: describe data collection strategy once finalised]

Current Status

This project is in early development. Data collection and model selection are ongoing. Results and publications will be added here as the work progresses. [TODO: update with progress]

Connection to the Broader PhD Project

Kurdish Voices is the spoken-language thread of the Dialect-Aware NLP PhD project, complementing the text-focused Dialectal Filtering work. Together they aim to cover both the written and spoken dimensions of Kurdish dialectal NLP.