English to Central Kurdish Speech Translation: Corpus Creation, Evaluation, and Orthographic Standardization
Authors: Mohammad Mohammadamini, Daban Q. Jaff, Josep Crego, Marie Tahon, Antoine Laurent
Organizations: LIUM, Le Mans University, Le Mans, France · Erfurt University, Erfurt, Germany · Koya University, Koysinjaq, Iraq · SYSTRAN (ChapsVision), Paris, France
We present KUTED, a speech-to-text translation (S2TT) dataset for Central Kurdish, derived from TED and TEDx talks. The corpus comprises 91,000 sentence pairs, including 170 hours of English audio, 1.65 million English tokens, and 1.40 million Central Kurdish tokens. We evaluate KUTED on the S2TT task and find that orthographic variation significantly degrades Kurdish translation performance, producing nonstandard outputs. To address this, we propose a systematic text standardization approach that yields substantial performance gains and more consistent translations. On a test set separated from TED talks, a fine-tuned Seamless model achieves 15.18 BLEU, and we improve Seamless baseline by 3.0 BLEU on the FLEURS benchmark. We also train a Transformer model from scratch and evaluate a cascaded system that combines Seamless (ASR) with NLLB (MT).
Figures & tables
Dataset
Source
Target
Size (hours)
Kurdish
Must-C
EN
14 langs
[237,505]
×
TEDx
EN
7 langs
[15,189]
×
Aug-LibriSpeech
EN
FR
212
×
CoVoST 2 EN → X
EN
15 langs
415
×
CoVoST 2 X → EN
21 langs
EN
[3,225]
×
VoxPopuli
15 European langs
15 European langs
[1,463]
×
Table 1: Speech translation datasets. The size column shows the range of speech data (hours) for language pairs in each dataset.
Split
TEDs
Utt
EN speech
EN tokens
CKB tokens
Clean
1,229
75k
138h
1.35m
1.14m
Noisy
467
16k
32h
0.30m
0.26m
All
1,696
91k
170h
1.65m
1.40m
Table 2: KUTED corpus specifications
Figure 1: The duration distribution of speech files in KUTED dataset
Evaluator
Accuracy
Fluency
Adaptation
Orthog.
Evaluator 1
4.68
4.59
4.63
4.73
Evaluator 2
4.30
4.22
4.16
4.41
Evaluator 3
3.83
3.80
3.76
4.03
Average
4.27
4.20
4.18
4.30
Table 3: Translation quality evaluation by professional translators.
Samples
Correct
S-shift
L-shift
Incorrect
Non-EN
Before filtering
1,000
925
53
16
7
2
ASR filtering
1,000
966
30
4
0
0
Table 4: Audio misalignment detection. S-shift = small shift; L-shift = large shift; Incorrect = audio and transcription not aligned; Non-EN = audio not in English.
Normalization
N0
N1
N2
N3
Unique Tokens
235,674
145,685
125,800
118,643
Table 5: Unique tokens after each step of normalization. N0 represents the original Kurdish transcriptions; N1–N3 are the three steps of normalization/standardization discussed in this section.
Figure 2: The Seamless V2 S2TT architecture
Figure 3: The cascade S2TT system
Metric
Train
Test
Validation
TEDs
1,668
16
12
Men
–
6
4
Women
–
6
4
Children
–
2
2
Noisy TEDs
463
2
2
Utterances
89,398
1,006
678
Table 6: Data partition for training, testing, and validation.
Parameter
Value
Learning rate
1e-4
Warmup steps
100
Patience
50
Batch size
6
Max epochs
10
Table 7: Seamless fine-tuning hyperparameters
System
KUTED
FLEURS
Seamless Baseline
5.04
9.36
Seamless KUTED
13.51
12.50
Table 8: E2E S2TT results for Seamless model on KUTED and FLEURS benchmarks.
System
N1
N2
N3
Seamless Baseline
4.71
5.04
5.47
Seamless KUTED
11.34
13.51
15.18
Table 9: Impact of standardization on CKB S2TT. The results are BLEU score evaluated on KUTED test set.
System
ASR Seamless (↓WER)
MT NLLB (↑BLEU)
Seamless–NLLB Baseline
21.62
9.25
Seamless–NLLB KUTED
8.62
15.57
Table 10: Cascade S2TT system results (lower WER is better; higher BLEU is better).