MGhana-ST: A Low-Resource Speech Translation Dataset for Ghanaian Languages and an Analysis of Multilingual Training Trade-offs

Frank Lawrence Nii Adoquaye Acquaye; Eric George Parakal; Jesse Johnson; Kishankumar Bhimani; Jochebed Afua Basil

cs.CLSep 30, 2026

MGhana-ST: A Low-Resource Speech Translation Dataset for Ghanaian Languages and an Analysis of Multilingual Training Trade-offs

Authors: Frank Lawrence Nii Adoquaye Acquaye, Eric George Parakal, Jesse Johnson, Kishankumar Bhimani, Jochebed Afua Basil

Organizations: Ashesi University, Berekuso, Ghana · AdwumaTech AI, Accra, Ghana · HSE University, Moscow, Russia · GIFT International Fintech Institute, Gandhinagar, India

Abstract

We present MGhana-ST, a speech translation dataset for four low-resource Ghanaian language varieties: Ga, Twi (Akuapem and Asante), Ewe, and Fante. MGhana-ST is an ongoing annotation effort; the experiments here use a fixed subset of about 16.1 hours of paired speech and English translations. The audio is curated from two existing Ghanaian speech resources. Unlike in those resources, the English translations are produced directly from audio by 37 native-speaker annotators and include verbal and non-verbal event annotations. Using Whisper-small, we compare monolingual and multilingual training under severe data scarcity, reporting means over three seeds. Flat multilingual training benefits no variety in this regime. Ga and Twi are unchanged within seed variance (+0.51 and +0.06 BLEU against monolingual standard deviations of 1.63 and 2.20), while Ewe declines by 6.99 BLEU and Fante by 5.11. The degrading varieties are Ewe, which is linguistically distinct and drawn from a different source corpus, and Fante, the least-resourced. Comparing empirical cross-lingual transfer with typology-based similarity, we find that transfer BLEU identifies closely interacting language pairs better than URIEL similarity, though neither predicts which varieties benefit from joint training. We also report a methodological finding. An earlier single-run analysis found positive transfer for three of four varieties; this did not survive replication across seeds. For Ga and Twi, monolingual baselines trained on 1.6 to 6.2 hours of audio have seed standard deviations roughly five and thirty times those of the multilingual models (0.35 and 0.07 BLEU). When the monolingual condition is noisier, a single-run comparison can show apparent transfer of this size from seed variation alone. We release MGhana-ST to support research on African language speech technology and low-resource speech translation.

Figures & tables

Language	Family	Train (h)	Val (h)	Test (h)
Twi	Akan	6.21	0.87	0.79
Ga	Ga-Dangme	2.48	0.29	0.27
Ewe	Gbe	2.47	0.37	0.34
Fante	Akan	1.62	0.19	0.18
Total		12.78	1.72	1.58

Table 1: The MGhana-ST experimental subset: train, validation, and test audio hours per language. This is the fixed subset used for all experiments reported here.

Language	Family	Train	Val	Test	Total	Total	Mean dur.
		utts.	utts.	utts.	utts.	(h)	(s)
Twi	Akan	11,849	1,593	1,484	14,926	7.87	1.90
Ga	Ga-Dangme	5,278	624	639	6,541	3.05	1.68
Ewe	Gbe	2,714	405	344	3,463	3.18	3.31
Fante	Akan	4,328	522	524	5,374	1.99	1.33

Table 2: Per-language corpus characteristics of the experimental subset. Utterance counts include all annotations; where a clip carries more than one independent translation, all references are retained. Total hours may differ from the sum of the per-split hours in Table 1 by 0.01 h due to rounding.

Language	Unique	Target	Multi-ref.
	refs.	overlap	rate
Twi	1,078	27.4%	2.7%
Ga	465	27.2%	0.6%
Ewe	339	1.5%	0.3%
Fante	323	38.4%	4.2%

Table 3: Reference diversity on the test split. Target overlap is the share of test utterances whose English reference is not unique (

1-\text{unique refs.}/\text{test utts.}

, with test utterance counts from Table 2 ). Ewe references are almost entirely unique, whereas over a third of Fante references are repeated.

Figure 1: MGhana-ST inference graph. Input speech is processed by the Whisper feature extractor and encoder, projected to the decoder dimension, and decoded autoregressively to produce an English translation.

Pair	URIEL	Geo.	Transfer
	sim. $\uparrow$	dist. $\downarrow$	BLEU $\uparrow$
Twi – Fante	0.772	0.109	18.15
Ga – Ewe	0.789	0.087	0.13
Ga – Twi	0.874	0.119	0.68
Twi – Ewe	0.794	0.165	0.14
Ga – Fante	0.767	0.211	1.23

Table 4: Pairwise language relatedness across complementary measures. Transfer BLEU is the mean of the two directional zero-shot scores for each pair, single-run at seed 42. URIEL identifies Ga–Twi as closest, while transfer BLEU identifies Twi–Fante as the strongest transfer pair.

Lang.	M1 (mono)	M2 (multi)	$\Delta$
BLEU
Ga	35.25 $\pm$ 1.63	35.76 $\pm$ 0.35	$+0.51$
Twi	18.79 $\pm$ 2.20	18.85 $\pm$ 0.07	$+0.06$
Ewe	15.48 $\pm$ 0.56	8.49 $\pm$ 1.15	$\bm{-6.99}$
Fante	47.61 $\pm$ 0.79	42.50 $\pm$ 1.43	$\bm{-5.11}$
chrF

Table 5: Monolingual (M1) versus flat multilingual (M2) performance per variety, mean

\pm

sample standard deviation over three training seeds with the data split held fixed. Bold marks deltas whose magnitude exceeds twice the larger of the two standard deviations. Corpus-level M2 BLEU is

22.50\pm 0.27

. Per-seed values are in Appendix F .

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Unfrozen layers ( $n$ )	BLEU	chrF
0	36.75	49.59
2	35.21	50.81
4 ∗	36.25	53.31
6	36.63	49.45
12	29.04	44.53

Appendix

Table 6: Encoder fine-tuning depth sweep on the Ga test split of the preliminary partition (single-run). ∗ Depth selected for main experiments.

Category	Ga	Twi	Ewe	Fante
Annotated segs.	19,370	37,100	11,333	22,780
Audio missing	0	12	4,516	6,917
Event-only	12,822	22,156	3,346	10,487
Other	7	3	3	0
Retained	6,541	14,929	3,468	5,376
Retention rate	33.8%	40.2%	30.6%	23.6%

Appendix

Table 7: Segment retention by exclusion category. Ten retained segments (Twi 3, Ewe 5, Fante 2) do not appear in the final splits, so the totals in Table 2 are slightly lower.

Model	Ga		Twi		Ewe		Fante
Model	BLEU	chrF	BLEU	chrF	BLEU	chrF	BLEU	chrF
M1-ga	34.76	49.25	0.98	8.25	0.07	7.13	1.48	9.24
M1-twi	0.38	8.11	17.91	29.70	0.17	9.21	23.36	43.17
M1-ewe	0.18	8.06	0.11	9.87	14.06	31.01	0.23	6.94
M1-fante	0.97	9.06	12.94	23.72	0.13	9.74	45.73	59.83

Appendix

Table 8: Cross-lingual transfer matrix: each M1 model is evaluated on all four language test sets (single-run, seed 42). Bold diagonal entries are in-language results. Off-diagonal entries are zero-shot transfers.

Run	Ga	Twi	Ewe	Fante	Corpus
BLEU
M2 (multi)	37.71	19.59	9.60	46.19	23.72
M3 (DANN)	37.74	18.76	13.23	41.62	23.38
M4 (LID)	36.12	19.08	8.96	44.43	22.96

Appendix

Table 9: Auxiliary conditioning results (single-run, seed 42). All comparisons are against M2 at the same seed, not against the three-seed mean in the main text.

Cfg	Seed	Ga	Twi	Ewe	Fante
BLEU
M1	1	36.83	17.25	16.10	46.82
M1	2	33.57	21.31	15.31	47.60
M1	3	35.35	17.80	15.02	48.40
M2	1	35.98	18.78	7.21	41.48
M2	2	35.35	18.92	9.44	44.13

Appendix

Table 10: Per-seed BLEU for M1 and M2, split held fixed. Corpus-level M2 BLEU is 22.30, 22.81, and 22.39 for seeds 1–3.

Explore similar work

Mar 14, 2026

GhanaNLP Parallel Corpora: Comprehensive Multilingual Resources for Low-Resource Ghanaian Languages

Low resource languages present unique challenges for natural language processing due to the limited availability of digitized and well structured linguistic data. To address this gap, the GhanaNLP initiative has developed and curated 41,513 parallel sentence pairs for the Twi, Fante, Ewe, Ga, and Kusaal languages, which are widely spoken across Ghana yet remain underrepresented in digital spaces. Each dataset consists of carefully aligned sentence pairs between a local language and English. The data were collected, translated, and annotated by human professionals and enriched with standard structural metadata to ensure consistency and usability. These corpora are designed to support research, educational, and commercial applications, including machine translation, speech technologies, and language preservation. This paper documents the dataset creation methodology, structure, intended use cases, and evaluation, as well as their deployment in real world applications such as the Khaya AI translation engine. Overall, this work contributes to broader efforts to democratize AI by enabling inclusive and accessible language technologies for African languages.

Lawrence Adu Gyamfi, Paul Azunre, Stephen Edward Moore +4

Ghana Natural Language Processing, Ghana · Khaya AI LTD, Ghana · Department of Mathematics, University of Cape Coast, Cape Coast, Ghana

Low-Resource Language Processing Low-Resource MT

Ask PDF

Apr 17, 2026

NaijaS2ST: A Multi-Accent Benchmark for Speech-to-Speech Translation in Low-Resource Nigerian Languages

Speech translation for low-resource languages remains fundamentally limited by the scarcity of high-quality, diverse parallel speech data, a challenge that is especially pronounced in African linguistic contexts. To address this, we introduce NaijaS2ST, a parallel speech translation dataset spanning Igbo, Hausa, Yorùbá, and Nigerian Pidgin paired with English. The dataset comprises approximately 50 hours of speech per language and captures substantial variation in speakers and accents, reflecting realistic multilingual and multi-accent conditions. With NaijaS2ST, we conduct a comprehensive benchmark of cascaded, end-to-end (E2E), and AudioLLM-based approaches across bidirectional translation settings. Our results show that audio LLMs with few-shot examples are more effective for speech-to-text translation than cascaded and end-to-end methods trained on fine-tuned data. However, for speech-to-speech translation, the cascaded and audio LLM paradigms yield comparable performance, indicating that there is still considerable room for improvement in developing targeted, task-specific models for this setting. By providing both a high-quality dataset and a systematic benchmark, we hope that NaijaS2ST will serve as a strong foundation for advancing research in low-resource, multilingual speech translation.

Marie Maltais, Yejin Jeon, Min Ma +7

Mila - Quebec AI Institute · McGill University, Canada · Google DeepMind +6

Speech Translation Audio-Language Models

Ask PDF

May 5, 2026

Nsanku: Evaluating Zero-Shot Translation Performance of LLMs for Ghanaian Languages

Large language models (LLMs) have demonstrated impressive multilingual capabilities for well-resourced languages, yet their performance on low-resource African languages remains poorly understood and largely unevaluated. This paper presents Nsanku, a systematic benchmark that evaluates the zero-shot machine translation performance of 19 open-weight and proprietary LLMs across 43 Ghanaian languages paired with English. Evaluation sentences were sourced from the YouVersion Bible platform, providing 300 sentence pairs per language. Two complementary automatic metrics are employed: Bilingual Evaluation Understudy (BLEU) and Character n-gram F-Score (chrF), alongside an average accuracy score and a cross-language consistency dimension. Nsanku represents the most comprehensive LLM translation evaluation for Ghanaian languages conducted to date. Results show that gemini-2.5-flash achieves the highest overall average score of 26.88 (BLEU: 24.60, chrF: 29.16), followed by claude-sonnet-4-5 at 24.87 (BLEU: 22.46, chrF: 27.28) and gpt-4.1 at 23.20 (BLEU: 21.15, chrF: 25.24). Among open-weight models, kimi-k2-instruct-0905 leads at an average score of 20.87. A critical finding from the consistency analysis is that no model and no language reached the Leaders quadrant of high performance and high consistency simultaneously, indicating that current LLMs are not yet reliably usable for Ghanaian language translation at scale. Siwu achieved the highest per-language average score at 25.73 while Nkonya scored lowest at 11.65. Nsanku establishes a publicly available, community-extensible evaluation infrastructure for African language NLP research.

Stephen E. Moore, Mich-Seth Owusu, Akwasi Asare +13

Department of Mathematics, University of Cape Coast, Cape Coast, Ghana · Ghana Natural Language Processing, Cape Coast, Cape Coast, Ghana · Khaya AI LTD Ghana

Multilingual Language Model Evaluation Open-Weight Language Models Low-Resource MT Low-Resource Language Processing Machine Translation MT Evaluation

Ask PDF