[Paper Review] Kurdish (Sorani) Speech to Text: Presenting an Experimental Dataset

Akam Qader, Hossein Hassani|arXiv (Cornell University)|Nov 29, 2019

Linguistics and Cultural Studies1 references4 citations

TL;DR

This paper introduces BD-4SK-ASR, the first experimental dataset for automatic speech recognition (ASR) in Sorani Kurdish, built using 200 core sentences from primary school textbooks (grades 1–3) in the Kurdistan Region of Iraq. The dataset, developed in the CMUSphinx framework, includes a 34-phone phoneme set, 2,000 synthesized sentences, transcriptions, narration files (16-bit, mono, 16 kHz), and a language model, enabling foundational ASR research for a low-resource language with no prior ASR resources.

ABSTRACT

We present an experimental dataset, Basic Dataset for Sorani Kurdish Automatic Speech Recognition (BD-4SK-ASR), which we used in the first attempt in developing an automatic speech recognition for Sorani Kurdish. The objective of the project was to develop a system that automatically could recognize simple sentences based on the vocabulary which is used in grades one to three of the primary schools in the Kurdistan Region of Iraq. We used CMUSphinx as our experimental environment. We developed a dataset to train the system. The dataset is publicly available for non-commercial use under the CC BY-NC-SA 4.0 license.

Motivation & Objective

• To address the critical lack of speech resources for Sorani Kurdish, a low-resource language with no prior ASR systems.
• To develop a foundational dataset for training and evaluating automatic speech recognition systems in Sorani Kurdish.
• To support initial experimentation in Kurdish ASR using a well-established, accessible framework like CMUSphinx.
• To lay the groundwork for future expansion to larger corpora and other Kurdish dialects.
• To promote open science by releasing the dataset under a CC BY-NC-SA 4.0 license for non-commercial research use.

Proposed method

• The dataset was constructed from 200 sentences extracted from Sorani Kurdish primary school textbooks (grades 1–3) in the Kurdistan Region of Iraq.
• A custom 34-phone phoneme set was defined to represent Sorani Kurdish sounds, aligned with Persian-Arabic script characters.
• 2,000 synthetic sentences were generated via random recombination of the 200 base sentences to increase word repetition and training coverage.
• Audio recordings were collected using Audacity in a quiet environment with a single speaker, at 16 kHz, 16-bit, mono format, with noise reduction applied.
• Transcriptions were created using the defined phoneme set and integrated into the CMUSphinx-compatible format for acoustic model training.
• A language model was built using CMUSphinx with fixed discounting (0.5) and ratio-based backoff, containing 283 unigrams, 5,337 bigrams, and 6,935 trigrams.

Experimental results

Research questions

• RQ1 Can a foundational ASR system be developed for Sorani Kurdish using a small, curated dataset from primary education materials?
• RQ2 What is the feasibility of using CMUSphinx for training an ASR system in a low-resource Kurdish language context?
• RQ3 How effective is a synthetic dataset of 2,000 sentences derived from 200 base sentences in supporting initial ASR training?
• RQ4 What phonetic and linguistic components are necessary to represent Sorani Kurdish accurately in an ASR pipeline?
• RQ5 How can open, non-commercial datasets accelerate research in under-resourced language technologies like Sorani Kurdish?

Key findings

• The BD-4SK-ASR dataset is publicly available under a CC BY-NC-SA 4.0 license, enabling non-commercial research and development in Sorani Kurdish ASR.
• The dataset includes a 34-phone phoneme set, transcriptions in phonetic format, and 2,000 audio files recorded at 16 kHz, 16-bit, mono, with minimal background noise.
• The language model contains 283 unigrams, 5,337 bigrams, and 6,935 trigrams, supporting n-gram language modeling for sentence-level recognition.
• The use of a single speaker for all recordings ensures consistency but limits generalization, highlighting a need for multi-speaker data in future work.
• The dataset serves as a first-step benchmark for Sorani Kurdish ASR, providing a foundation for future expansion to larger, more diverse corpora.
• The authors confirm the absence of prior literature on Kurdish ASR, establishing this work as the first experimental attempt in the field.