💬 NLP

English to Central Kurdish Speech Translation: Corpus Creation, Evaluation, and Orthographic Standardization

This paper introduces KUTED, a large-scale English-to-Central Kurdish speech translation corpus derived from TED talks, and demonstrates that implementing a systematic orthographic standardization approach significantly improves translation performance for both fine-tuned Seamless and cascaded Transformer models.

Original authors: Mohammad Mohammadamini, Daban Q. Jaff, Josep Crego, Marie Tahon, Antoine Laurent

Published 2026-04-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohammad Mohammadamini, Daban Q. Jaff, Josep Crego, Marie Tahon, Antoine Laurent

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a library of 170 hours of fascinating TED Talks, but they are all in English. Now, imagine you want to share these ideas with millions of people who speak Central Kurdish, a beautiful language spoken by about 8 million people across Iraq, Iran, Turkey, and Syria.

The problem? There is almost no "dictionary" or "training manual" (data) that teaches computers how to listen to English speech and instantly translate it into Kurdish text. Without this data, AI is like a student trying to learn a language without a textbook.

This paper is about building that textbook from scratch. Here is the story of how they did it, explained simply:

1. The Big Project: "KUTED" (The Kurdish TED Library)

The researchers created a massive new dataset called KUTED. Think of this as a giant, synchronized library.

The Source: They took 1,696 TED and TEDx talks.
The Team: They didn't just use robots. They partnered with a community of Kurdish volunteers (mostly university students) who manually translated these talks.
The Result: They now have 170 hours of English audio, paired perfectly with English text and Kurdish text. It's like having a movie with the audio, the English subtitles, and the Kurdish subtitles all lined up perfectly.

2. The "Messy Handwriting" Problem

Here is the tricky part. Kurdish is a bit like a language where people write the same word in many different ways.

The Analogy: Imagine if some people wrote "color" and others wrote "colour," but also some wrote "colr" or "k-l-r." If you were trying to learn to spell, you would get confused.
The Reality: In Kurdish, the same word might be written as one long block, or split into three separate pieces, or spelled with different letters depending on who typed it. This "messy handwriting" (orthographic variation) confuses AI. The AI thinks "word A" and "word B" are totally different words, even though they mean the same thing. This makes the AI's translation terrible.

The Solution: The researchers acted like a strict but helpful editor. They created a "Standardization Rulebook." They went through the entire library and forced every word to be written in the same, correct way.

The Result: After cleaning up the "messy handwriting," the AI's performance jumped significantly. It's like taking a blurry, pixelated photo and sharpening it until the picture is crystal clear.

3. Teaching the AI (The Three Methods)

Once they had the clean data, they tried three different ways to teach the AI to translate:

Method A: The "Fine-Tuning" Approach (The Smart Tutor)
They took a very smart, pre-trained AI (called Seamless) that already knew many languages. They gave it the KUTED library to study.
- Result: The AI got much smarter. It went from knowing almost nothing about Kurdish to translating with decent accuracy. It even got better at translating other types of Kurdish text it hadn't seen before.
Method B: The "Assembly Line" (The Two-Step Process)
Instead of one robot doing everything, they used two robots working in a line:
1. Robot 1 (ASR): Listens to the English audio and types it out.
2. Robot 2 (MT): Takes that typed English and translates it to Kurdish.
- Result: This worked almost as well as the single smart robot, proving that if you have a good listener and a good translator, the team works great.
Method C: The "From Scratch" Approach (The Blank Slate)
They tried to build a translator from zero, using only the KUTED library, without any help from pre-trained models.
- Result: It struggled a bit. This taught them that while KUTED is amazing, it's not quite big enough to teach a robot everything on its own. It needs to be combined with other data sources to be perfect.

4. Why This Matters

Before this paper, Kurdish was like a "ghost town" for speech translation technology. There were very few resources.

The Impact: This paper provides the first major "training ground" for Kurdish speech translation.
The Lesson: It proved that cleaning up the data (fixing the spelling) is just as important as the AI model itself. If you feed a smart AI messy data, it gives messy answers. If you give it clean, standardized data, it shines.

In a Nutshell

The researchers built a massive, high-quality library of English-to-Kurdish translations. They realized the Kurdish text was written in a confusing, inconsistent way, so they cleaned it up like a librarian organizing a messy bookshelf. Then, they used this clean library to teach AI how to translate English speech into Kurdish. The result? A huge leap forward for Kurdish technology, showing that with the right data and a little bit of organization, even low-resource languages can get a seat at the high-tech table.

1. Problem Statement

The paper addresses two critical bottlenecks in Speech-to-Text Translation (S2TT) for Central Kurdish (CKB):

Data Scarcity: There is a lack of large-scale, high-quality parallel speech-text datasets for Kurdish, particularly for the English-to-Kurdish direction. Existing resources are either small, pseudo-labeled, or limited to text-only translation.
Orthographic Variability: Central Kurdish lacks a fully standardized writing system. Significant variation exists in spelling, word segmentation (joined vs. non-joined words), and the use of loanwords/proper names. This "orthographic noise" increases data sparsity and severely degrades the performance of Neural Machine Translation (NMT) and S2TT models, as they struggle to generalize across different surface forms of the same token.

2. Methodology

A. Corpus Creation: KUTED

The authors introduce KUTED (Kurdish TED), a new S2TT dataset derived from TED and TEDx talks.

Source: 1,696 verified TED/TEDx talks with human-translated Kurdish subtitles.
Scale: 170 hours of English audio, 91,000 sentence pairs, 1.65 million English tokens, and 1.40 million CKB tokens.
Data Collection & Cleaning:
- Audio was obtained from TED/YouTube and converted to 16kHz mono.
- Talks were categorized as "clean" (1,229) or "noisy" (467, containing background noise).
- Alignment: A four-step process was used: TED-level realignment, sub-sentence realignment (handling offsets <1s), sentence-level alignment, and audio extraction.
- Quality Control: A human evaluation of 1,000 samples showed 92.2% correct audio alignment.
- Automated Filtering: An ASR-based method (using Seamless v2) was developed to detect misaligned audio by computing the normalized Levenshtein distance between the ASR hypothesis and the reference transcript. Samples with a distance $>0.3$ were filtered, removing ~4,000 misaligned instances.

B. Orthographic Standardization

To combat orthographic variability, the authors proposed a systematic three-step standardization pipeline:

Normalization (N1): Used the AsoSoft text normalizer for Unicode correction, punctuation standardization, and number unification. Punctuation was separated from tokens to reduce lexicon size.
General Correction (N2): Applied a correction table of 19,700 pairs derived from the AsoSoft corpus (the largest available Kurdish text corpus) to fix frequent misspellings.
KUTED-Specific Correction (N3): Extracted unique tokens from the normalized KUTED corpus and manually reviewed high-frequency variants to create a specific correction table (11,860 pairs). This step replaced ~150,000 token instances (10% of the corpus).

Impact: This process reduced the number of unique tokens by approximately 50%, significantly densifying the data distribution.

C. Experimental Setup

The authors evaluated the dataset using three distinct system architectures:

End-to-End (E2E) S2TT: Fine-tuning the Seamless model (W2V-BERT 2.0 encoder + NLLB decoder) on KUTED.
Cascaded S2TT: A pipeline where a fine-tuned Seamless model performs ASR, followed by a fine-tuned NLLB-1.3B model for Machine Translation (MT).
From-Scratch Training: Training a Fairseq Transformer model (76M parameters) solely on KUTED (speech encoder pretrained on English ASR only; decoder unpretrained) to test corpus quality without relying on massive pretraining.
Text-to-Text (T2TT): Fine-tuning NLLB for EN↔CKB translation to establish a baseline for text-only performance.

3. Key Contributions

KUTED Dataset: The first large-scale (170 hours), human-aligned English-to-Central Kurdish speech translation corpus.
Standardization Framework: A reproducible, three-step orthographic standardization methodology that addresses the specific morphological and orthographic challenges of Central Kurdish.
Misalignment Detection: A robust, ASR-based method for automatically detecting and filtering misaligned audio-transcript pairs.
Comprehensive Evaluation: Extensive benchmarking across E2E, cascaded, and from-scratch models, including out-of-domain testing on the FLEURS benchmark.

4. Results

Performance Gains

Orthographic Impact: Standardization was crucial. For the fine-tuned Seamless model, applying the full standardization pipeline (N1+N2+N3) increased the BLEU score on the KUTED test set from 11.34 (N1 only) to 15.18. The baseline (pretrained, unstandardized) showed negligible gains from standardization alone, confirming that the model must be trained on standardized data to benefit.
E2E S2TT (Seamless):
- KUTED Test: Improved from a baseline of 5.04 BLEU to 13.51 BLEU after fine-tuning on KUTED.
- FLEURS (Out-of-Domain): Improved from 9.36 BLEU to 12.50 BLEU (+3.0 BLEU), demonstrating strong generalization.
Cascaded System:
- Fine-tuning the ASR component reduced Word Error Rate (WER) from 21.62% to 8.62%.
- The cascaded system (ASR + MT) achieved 15.57 BLEU, slightly outperforming the E2E Seamless model (13.51 BLEU).
From-Scratch Model: The Fairseq model trained from scratch achieved 7.90 BLEU. While lower than fine-tuned models, this result validates that KUTED contains sufficient signal for learning, though pretraining is necessary for high performance given the dataset size.
Text-to-Text (T2TT): The fine-tuned NLLB model achieved 16.72 BLEU (EN→CKB) and 27.93 BLEU (CKB→EN), establishing a strong text-only baseline.

5. Significance

Resource for Low-Resource Languages: KUTED fills a critical gap for Central Kurdish, a language spoken by ~8 million people but lacking large-scale speech resources.
Proof of Orthographic Necessity: The study provides empirical evidence that for agglutinative languages with non-standardized orthography (like Kurdish), text standardization is not optional but a prerequisite for achieving competitive S2TT performance.
Community Mobilization: The paper documents a successful workflow involving the "Kurdish TED Translator Community" and university volunteers, offering a blueprint for creating similar resources for other low-resource languages.
Benchmarking: The results on FLEURS demonstrate that training on high-quality, standardized in-domain data (KUTED) significantly improves performance on out-of-domain tasks, validating the dataset's utility for general speech translation research.

Limitations & Future Work:
The authors note that the dataset cannot be publicly shared due to TED copyright restrictions; only the list of Talk IDs is available. Future work aims to extend this methodology to other low-resource languages with existing TED translations and to automate the script standardization process using learned models.