DonorRank: Donor Language Selection for Low-Resource Cross-Lingual Speech Recognition

# DonorRank: Donor Language Selection for Low-Resource Cross-Lingual Speech Recognition
Source: [https://arxiv.org/html/2608.11441](https://arxiv.org/html/2608.11441)
\[ Script=Devanagari, Scale=MatchLowercase, Path=\./, \]

Aarohi SrivastavaDavid ChiangAffiliation:Computer Science and EngineeringAffiliation:University of Notre DameAffiliation:Notre Dame, IN, USAAffiliation:\{adhasman, asrivas2, dchiang\} @nd\.edu

###### Abstract

Low\-resource automatic speech recognition \(ASR\) commonly relies on cross\-lingual transfer, where models are adapted from higher\-resource donor languages\. However, selecting donors remains challenging for spontaneous speech from under\-resourced language communities, due to linguistic variation, evolving orthographic conventions, and uneven resource availability\. We present DonorRank, a learning\-to\-rank framework for predicting effective donor languages for zero\-shot ASR\. We evaluate DonorRank on two multilingual speech corpora of Indic and African language families\. It accurately predicts donor language rankings and improves donor selection over common heuristics based on genetic similarity or high\-resource languages\. Beyond improving transfer, we show how DonorRank is a general framework for analyzing donor language selection itself\. Our analyses show that the composition of the donor set determines which linguistic cues are useful in predicting successful transfer\. We also identify transfer patterns that provide practical guidance for multilingual ASR in low\-resource settings\.

## 1Introduction

Building automatic speech recognition \(ASR\) systems for the world’s thousands of languages is challenging since most languages lack sufficient transcribed speech for supervised training\. Cross\-lingual transfer has thus become a common strategy for low\-resource ASR, where a model is fine\-tuned on one or more higher\-resource donor languages before being applied to an unseen target language\. However, the transfer performance in this approach depends critically on the choice of donor languages, and can vary substantially even among closely related languages\. Therefore, what makes a donor language effective is an important question for developing robust and inclusive multilingual speech technologies\.

Cross\-lingual transfer in low\-resource ASR presents additional challenges beyond those typically considered in multilingual language processing\. Unlike many benchmarks built from standardized text or read speech, low\-resource ASR often relies on spontaneous speech collected in real\-world settings, where pronunciation, speaking style and recording conditions exhibit substantial variability\. These challenges are compounded by the fact that many low\-resource languages have limited orthographic standardization making lexical similarity less reliable\. Consequently, identifying effective donor languages requires accounting for multiple interacting linguistic and dataset\-specific factors, rather than relying on genealogical relatedness or lexical similarity alone\.

To better understand how these factors impact donor language selection, we examine two multilingual collections that represent markedly different low\-resource ASR scenarios\. The first is a subset of the VAANI corpus[20](https://arxiv.org/html/2608.11441#bib.bib6)containing Devanagari\-script Indic language varieties \(which we call VAANI\-D\), where transfer occurs among closely related languages sharing a common writing system\. The second uses the WAXAL corpus[7](https://arxiv.org/html/2608.11441#bib.bib16), which spans African languages from multiple language families and writing systems, representing a substantially more heterogeneous multilingual setting\. Studying these complementary collections allows us to examine how the linguistic signals governing successful transfer change between closely related and typologically diverse language ecosystems\.

Our study makes the following contributions:

1. 1\.We presentDonorRank, a learning\-based framework for donor language selection in zero\-shot ASR and evaluate it in two complementary multilingual settings\.
2. 2\.We provide an empirical analysis of the features that drive donor language selection, demonstrating that linguistic and dataset characteristics improve donor rankings over those based solely on genealogical relationships\.
3. 3\.We identify recurring transfer patterns, including effective donor hubs and non\-obvious donor languages, providing practical guidance for donor language selection in low\-resource ASR scenarios where empirical evidence is limited\.

## 2Related Work

#### Donor Language Selection in Text

Selecting donor languages for transfer learning in low\-resource settings has traditionally relied on heuristics based on genealogical relatedness or linguistic intuition\. To move beyond such methods,[14](https://arxiv.org/html/2608.11441#bib.bib2)introduced LangRank, a learning\-based framework fortext processingthat predicts transfer performance using typological, inventory, and geographic features from the URIEL/lang2vec database[15](https://arxiv.org/html/2608.11441#bib.bib7)along with features specific to the dataset\. Similarly, NN\-Rank[9](https://arxiv.org/html/2608.11441#bib.bib3)uses learned dataset embeddings for ranking source languages in text\-based NLP tasks\. Subsequent text processing work has examined how linguistic features influence transfer[18](https://arxiv.org/html/2608.11441#bib.bib8)\. These approaches have been studied in text\-based NLP tasks, but they are promising candidates for the low\-resource ASR setting\. At the same time, extending donor selection to speech processing is non\-trivial, as it depends not only on linguistic similarity but also on acoustic variability and dataset\-specific factors\.

#### Donor Language Selection in Speech

In speech recognition, prior work has explored donor selection using data\-driven similarity measures\. These include methods based on acoustic token distributions[19](https://arxiv.org/html/2608.11441#bib.bib5);[16](https://arxiv.org/html/2608.11441#bib.bib4), as well as approaches that leverage spoken language identification models to filter training data[2](https://arxiv.org/html/2608.11441#bib.bib11)\. These approaches have shown to be more effective in predicting the downstream ASR performance than relying only on linguistic features[19](https://arxiv.org/html/2608.11441#bib.bib5); however, they typically require processing large amounts of audio data and offer limited interpretability\. While feature\-based similarity has been used for donor selection in ASR tasks, we present a LangRank\-inspired learning\-to\-rank framework for donor selection for low\-resource languages and evaluate and analyze it on several nonstandard and minority language varieties from multiple African and Indic language families\.

#### Linguistic Predictors of Transfer

While dataset features such as the number of hours of training data and the number of uniquely occurring words play a vital role in determining cross\-lingual transfer success, recent work[6](https://arxiv.org/html/2608.11441#bib.bib17);[1](https://arxiv.org/html/2608.11441#bib.bib18)has explored the impact of linguistic features such as lexical\-phonetic distances, language family belonging, and shared writing systems on the success of cross\-lingual transfer for several text\-based NLP tasks\. However, such in\-depth studies have not been explored on a large scale for ASR\.[10](https://arxiv.org/html/2608.11441#bib.bib19)pursue this direction in ASR, exploring the impact of feature similarity in cross\-lingual transfer for ASR within eight African languages\.

#### Our Work

In contrast, we focus on zero\-shot ASR transfer to low\-resource language varieties and dialects, where data is often sparse, noisy, and drawn from spontaneous speech\. Our study focuses on Indo\-Aryan, Niger\-Congo, Afro\-Asiatic, and Nilo\-Saharan languages that are linguistically diverse and underrepresented in ASR and NLP resources\. Despite the large number of speakers, many of these varieties are absent from standard ASR benchmarks, making them a compelling testbed for evaluating transfer methods in realistic low\-resource settings\. Inspired by LangRank[14](https://arxiv.org/html/2608.11441#bib.bib2), our approach,Donor\-Rank, combines linguistic and dataset\-level features in a lightweight ranking framework, enabling both efficient donor selection and interpretable analysis of factors influencing transfer\.

## 3Data and Languages

We evaluate donor language selection on two spontaneous speech corpora representing contrasting multilingual transfer settings: VAANI\-D, a Devanagari\-script subset of the VAANI corpus[20](https://arxiv.org/html/2608.11441#bib.bib6), and WAXAL[7](https://arxiv.org/html/2608.11441#bib.bib16)\. VAANI\-D consists of closely related Indo\-Aryan language varieties spoken across northern and central India, providing a linguistically controlled setting for studying transfer within a dense network of related languages\. In contrast, WAXAL spans languages from multiple African language families spoken across Sub\-Saharan Africa and written in multiple scripts, providing a complementary setting for studying donor selection across substantially greater typological diversity\. These datasets allow us to examine donor language selection in two distinct low\-resource ASR regimes: transfer among closely related language varieties and transfer across broader multilingual landscapes\.

#### VAANI\-D

VAANI is a large\-scale corpus of spontaneous speech collected by prompting speakers to describe images in their local language or dialect\. The full dataset contains over 150,000 hours of speech, approximately 10% of which is transcribed, from more than 156,000 speakers across 773 districts of India\. We restrict our experiments to a Devanagari\-script subset of VAANI, referred to asVAANI\-D\. This subset comprises 20 Indic languages and language varieties, including both higher\- and lower\-resource varieties\. Among these, Awadhi \(awa\), Bhili \(bhb\), Garhwali \(gbm\), Halbi \(hlb\), Konkani \(kok\), and Marwari \(mwr\) serve as the primary low\-resource target languages for evaluating zero\-shot transfer\.

Restricting our analysis to a common writing system provides a linguistically controlled setting in which differences in transfer performance are less likely to arise from script mismatch and instead reflect variation among closely related language varieties\. The corpus has spontaneous code\-mixed speech, and several language varieties exhibit limited orthographic standardization despite sharing the Devanagari script\. For instance, the VAANI\-D dataset contains Hindi which is a standardized language as well as several of its dialects like Awadhi, Khariboli, and Bundeli, which use micro\-variations in orthography and pronunciation\. Donor language selection in this setting can’t be explained solely by script or lexical similarity, making VAANI\-D an informative setting for studying transfer among closely related languages\.

#### WAXAL

To contrast with VAANI\-D, we evaluate on the ASR component of the WAXAL dataset, which contains approximately 1,250 hours of transcribed spontaneous speech across 19 African languages spoken by more than 100 million speakers\. These languages are typologically diverse, spanning multiple language families \(e\.g\., Bantu, Simetic\) and written scripts \(e\.g\., Latin, Ge’ez\)\. While VAANI\-D only includes Indo\-Aryan languages spoken in a specific region, WAXAL encompasses a much more diverse spectrum of highly distinct language families spanning a much larger geographic area\. This allows us to consider agglutinative languages such as Luganda \(lug\) and Lingala \(lin\) along with tonal languages such as Ewe, Ikposo \(kpo\) and Acholi \(ach\) as cross\-lingual donors\. Thus, WAXAL serves as a strong exploratory dataset for ranking across different language families\.

Many languages in WAXAL are represented using practical Romanization rather than long\-established written conventions, reflecting another common characteristic of low\-resource ASR\. WAXAL provides a complementary multilingual setting in which donor language selection must generalize beyond closely related language varieties\. Comparing results between VAANI\-D and WAXAL therefore allows us to examine how donor selection behaves under two different conditions\. We expect the donor\-selection to be more reliant on fine\-grained linguistic features for closely related languages in VAANI\-D\. In contrast we expect the donor selection to be dependent on broader orthographic and genetic similarities for the much more diverse WAXAL dataset\.

#### Data Setup

For both datasets, we fine\-tune the ASR model using the training split of a single donor language and evaluate zero\-shot transfer on unseen target languages using the corresponding test splits\. Development data is used for model selection\. To facilitate fair comparison across donor languages, we cap the amount of fine\-tuning data at seven hours per language and use one hour of test data per language\. Some languages in VAANI\-D contribute fewer than seven hours due to data availability; the exact amount of training data used for each language is reported in Table[8](https://arxiv.org/html/2608.11441#A1.T8)\.

## 4DonorRank

Figure 1:Illustration of DonorRank\. We assemble feature\-based language vectors using a composition of features for donor languagesdi∈Dd\_\{i\}\\in Dand target languagettto input to the ranker, which is supervised upon transcription error rate scores produced byw2vBERT\.Figure[1](https://arxiv.org/html/2608.11441#S4.F1)illustrates the DonorRank framework which extends the LangRank framework[14](https://arxiv.org/html/2608.11441#bib.bib2)\.

### 4\.1Learning Donor Rankings

We first construct ground\-truth donor rankings through pairwise transfer experiments\. For every donor\-target language pair, we fine\-tune a multilingual ASR model on speech from the donor language and evaluate it in a zero\-shot setting on the target language\. We choose w2vBERT[3](https://arxiv.org/html/2608.11441#bib.bib9)on the basis of preliminary experiments and related work[11](https://arxiv.org/html/2608.11441#bib.bib15)\. The resulting transcription error rates provide an empirical measure of transfer quality, allowing donor languages to be ranked independently for each target language\. These initial experiments, though somewhat compute\-intensive, are valuable in training the ranking framework which in turn preserves the intensive and costly experimentation currently required for effective donor selection for a plethora of languages in the zero\-shot low\-resource setting\.

Using these empirical rankings as supervision, we follow[14](https://arxiv.org/html/2608.11441#bib.bib2)in training a LightGBM[13](https://arxiv.org/html/2608.11441#bib.bib1)learning\-to\-rank model with the LambdaRank objective\. Given a target language and a set of candidate donors, the model predicts an ordering that maximizes agreement with the observed transfer rankings on the basis of linguistic and dataset feature vectors for the source and target language\. The DonorRank framework helps us select the optimal donor language based on the observed trends rather than directly improving the underlying ASR model\.

### 4\.2Linguistic and Dataset Features

Each donor\-target pair is represented using a combination of linguistic and dataset\-level features, illustrated in Figure[1](https://arxiv.org/html/2608.11441#S4.F1)\. This representation enables the ranking model to capture various sources of information that may influence cross\-lingual transfer\. The linguistic features are obtained from URIEL/lang2vec[15](https://arxiv.org/html/2608.11441#bib.bib7)and include genetic, geographic, syntactic, and phonological representations derived from WALS[8](https://arxiv.org/html/2608.11441#bib.bib12), SSWL[4](https://arxiv.org/html/2608.11441#bib.bib13), and PHOIBLE[17](https://arxiv.org/html/2608.11441#bib.bib14)\. These features encode different notions of language similarity, ranging from genealogical relationships to structural and phonological properties\. In addition, we calculate and incorporate dataset\-specific features: donor training hours, target training and test hours, target type\-to\-token ratio, and lexical overlap between donor and target transcriptions\. Depending on the composition of languages considered in donor selection ranking, different features are expected to be more or less relevant, which we explore in our analyses\.

### 4\.3Optimal Donor Prediction

At inference time, DonorRank predicts a ranked list of donor languages for a new target language, given concatenated linguistic and dataset features for the language\. The highest\-ranked donors are then selected for ASR fine\-tuning and subsequent zero\-shot evaluation\. We evaluate DonorRank using leave\-one\-target\-language\-out cross\-validation\. For each fold, donor\-target pairs involving the held\-out target language are excluded from training, and the ranking model predicts donor rankings for the unseen target language\. This evaluation protocol measures the practical setting of interest: selecting donor languages for a new low\-resource target language without prior knowledge of its transfer performance\. Because the ranking model is never trained on transfer results involving the evaluation language, successful ranking demonstrates generalization beyond previously observed language pairs\.

## 5Experimental Setup

Our experiments address four questions:

1. 1\.Can DonorRank accurately predict effective donor languages?
2. 2\.Do improved donor rankings improve zero\-shot ASR performance?
3. 3\.What linguistic and resource\-related factors relate to successful donor selection?
4. 4\.Does combining multiple highly\-ranked donor languages improve transfer?

#### General Setup

All experiments use the same pre\-trained w2vBERT checkpoint111[https://huggingface\.co/facebook/w2v\-bert\-2\.0](https://huggingface.co/facebook/w2v-bert-2.0)[5](https://arxiv.org/html/2608.11441#bib.bib20)and identical fine\-tuning hyperparameters\. Models are fine\-tuned for 10 epochs with a batch size of 16 and a learning rate of5×10−55\\times 10^\{\-5\}\.

#### Ranking Evaluation

We evaluate ranking fidelity using Normalized Discounted Cumulative Gain \(NDCG\), which measures agreement between predicted donor rankings and the empirical rankings obtained from pairwise transfer experiments[12](https://arxiv.org/html/2608.11441#bib.bib10)\.

#### Zero\-Shot Transfer Evaluation

To evaluate the scores obtained by the selected donor, we fine\-tune w2vBERT on the highest\-ranked donor language and evaluate zero\-shot ASR on the target language\. Performance is measured using character error rate \(CER\) and word error rate \(WER\) using Levenshtein distance between predicted and reference transcriptions\. We further compare DonorRank against donor selection based on two baselines: the closest phylogenetic neighbor for each target and a prominent high\-resource language in each collection\.

#### Multi\-Donor Transfer

Finally, we investigate whether combining multiple highly ranked donors further improves zero\-shot transfer\. Because these experiments require substantially more ASR training runs, we perform them on six primary low\-resource target languages in VAANI\-D: Awadhi, Bhili, Garhwali, Halbi, Konkani, and Marwari\. For each target language, we fine\-tune models using the top\-kkranked donor languages \(k=1,…,5k=1,\\ldots,5\), distributing the available fine\-tuning data approximately evenly across selected donors while maintaining a comparable training budget\.

## 6Results

We evaluate DonorRank’s ability to predict effective donor languages, then examine whether improved rankings translate into downstream ASR gains\. We also analyze the linguistic factors for successful transfer before concluding with a study of multi\-donor transfer\.

Table 1:Mean NDCG Scores for WER\-based and CER\-based DonorRanker on WAXAL and VAANI\-D\.### 6\.1DonorRank accurately predicts optimal donor languages\. Ranker metric choice depends on the data\.

Across both VAANI\-D and WAXAL, DonorRank achieves high agreement between predicted and empirical donor rankings, showing that donor selection can be learned from linguistic and resource\-related features \(see Table[1](https://arxiv.org/html/2608.11441#S6.T1)\)\. We train each ranker on either CER or WER scores and find that the choice of metric matters for ranking fidelity\.

For VAANI\-D, the CER\-based ranker has a higher mean NDCG \(0\.984\) than the WER\-based ranker\. A pairedtt\-test across target languages shows that this difference is statistically significant \(t⁡\(17\)=4\.02t\(17\)=4\.02,p<0\.001p<0\.001\)\. This shows that CER is a more reliable signal for ranking in VAANI\-D than WER\. VAANI\-D includes languages where spellings and word boundaries are not standardized, so character\-level differences may better capture cross\-variety similarity than word\-level measures\.

In contrast, the WER\-based ranker performs better on WAXAL \(0\.948 versus 0\.924\), but the difference is not significant\. WAXAL includes languages from different families and scripts with limited or no linguistic or phylogenetic relationship, and character\-level information may be too fine\-grained given the composition of the languages\. We see this reflected in how Amharic \(amh\) and Tigrinya \(tir\), the only two Semitic languages in WAXAL and sharing the Ge’ez script, are the best performing donors for each other \(see Table[3](https://arxiv.org/html/2608.11441#S6.T3)\)\. Although both ranking objectives produce highly accurate rankings \(\>0\.90 NDCG\), this difference suggests that the most informative signal may depend on the characteristics of dataset\. We use the ranker with a higher NDCG \(CER\-based for VAANI\-D and WER\-based for WAXAL\) for subsequent experiments\.

### 6\.2Accurate donor ranking improves zero\-shot transfer\.

Table 2:Performance comparison: Fine\-tuning w2vBERT on \(A\) DonorRank top\-1 donor, \(B\) top\-1 nearest genetically\-similar donor, and \(C\) Hindi\. Bold indicates the lowest \(best\) score \(%\)\. DonorRank consistently provides the best donor option\.DonorRank \(A\)Genetic \(B\)Hindi \(C\)TargetDonorCERWERDonorCERWERCERWERanpmai22\.153\.1mai22\.153\.137\.781\.1awahne22\.455\.0hne22\.455\.034\.381\.2bgcraj23\.257\.8hin32\.270\.532\.270\.5bhbkok41\.873\.5hne48\.491\.251\.694\.1bhomag19\.047\.2sdr42\.287\.230\.071\.4gbmkfy35\.264\.8kfy35\.264\.840\.186\.5hinmwr15\.634\.1bgc99\.910020\.446\.4hlbhne37\.266\.2kok44\.083\.445\.989\.4hnebho23\.155\.9awa99\.810032\.376\.1kfybho13\.333\.7gbm22\.351\.918\.545\.7kokmar31\.167\.7mar31\.167\.744\.797\.3magbho24\.454\.4anp25\.758\.034\.472\.3maimag20\.546\.8anp21\.849\.832\.871\.8marmjn27\.269\.7kok28\.468\.044\.098\.3mwrraj24\.147\.2bhb10010037\.071\.8nepkfy32\.0102kfy32\.010243\.9111rajmwr21\.546\.4bhb10010035\.973\.5sjpanp29\.869\.6hlb38\.789\.442\.393\.2

Table 3:Performance comparison: Fine\-tuning w2vBERT on \(A\) DonorRank top\-1 donor, \(B\) top\-1 nearest genetically\-similar donor, and \(C\) Oromo\. Bold indicates the lowest \(best\) score \(%\) for each target language\.Figure 2:Performance of w2vBERT fine\-tuned on top\-k ranked donors for multiple target varieties\. We find a negative correlation between CER andkk\(mean Pearson’sρ=−0\.531\\rho=\-0\.531\)\.We compare DonorRank against two practical donor selection baselines: selecting the nearest phylogenetically related language and selecting a single high\-resource baseline language\. We use Hindi as the high\-resource baseline for VAANI\-D and Oromo for WAXAL\. Both are the most widely spoken in their respective dataset\. DonorRank consistently picks the optimal donor across all 19 varieties in VAANI\-D compared to the genetic neighbor and Hindi baselines \(Table[2](https://arxiv.org/html/2608.11441#S6.T2)\)\. Whereas in WAXAL, we see that the genetically similar donor language performs at par or better with DonorRank’s top\-1 language more often than it does for VAANI\. Nevertheless, DonorRank often identifies donor languages that outperform both the genetic baseline and Oromo, demonstrating that learned ranking remains valuable even in a substantially more diverse multilingual setting

These results suggest that the role of donor ranking differs based on the composition of languages being considered\. In VAANI\-D, the challenge lies in distinguishing among many closely related candidate donors, where learned ranking consistently identifies the strongest transfer language\. In WAXAL, donor selection involves navigating a much broader multilingual landscape, where language relatedness provides a useful starting point but learned ranking can uncover more effective transfer partners\. At the same time, even the most effective donor can yield high error rates in some cases; while donor selection is a crucial starting point and the focus of this work, low\-resource ASR has a lot of room for improvement before systems compare to more dominantly studied languages \(e\.g\., English\)\.

### 6\.3Saturation Effect in Multiple Donors

In Figure[2](https://arxiv.org/html/2608.11441#S6.F2), we see that across all six target languages, incorporating multiple donors generally improves performance over single\-donor transfer\. We find that the highest\-ranked languages tend to capture most of the transferable information \(k=2​…​4k=2\.\.\.4\), while additional donors provide diminishing returns\. This suggests a saturation effect: combining a small number of well\-chosen donors is beneficial, but adding more languages yields limited gains and can introduce noise \(as in the case of Bhili\)\. These results show that DonorRank is also useful for constructing effective donor sets for multilingual adaptation\. These findings have practical implications for low\-resource ASR and dialectal settings, where variation can make indiscriminate aggregation of data harmful\.

## 7Understanding Donor Language Selection

DonorRank also provides a common framework for analyzing the factors governing successful transfer in different multilingual settings\.

![Refer to caption](https://arxiv.org/html/2608.11441v1/aggregate_feature_importance_7cats.png)Figure 3:Normalized feature importance for DonorRank on WAXAL and VAANI\-D\.### 7\.1What linguistic signals govern successful transfer?

Figure[3](https://arxiv.org/html/2608.11441#S7.F3)compares feature importance across VAANI\-D and WAXAL\. In VAANI\-D, the ranking model relies on geographic proximity, lexical overlap, and the amount of available training data\. These cues help distinguish among languages that are already closely related and some of which are mutually intelligible\.

In contrast, WAXAL places more emphasis on syntactic features while geographic information remains informative\. Because WAXAL spans multiple language families and writing systems, structural linguistic properties are stronger indicators of transferability than lexical similarity alone\. This suggests that the same ranking framework adapts to the linguistic composition of each multilingual collection, learning whichever signals best predict successful transfer\. Notably, genetic similarity does not emerge as an important feature for either dataset, with normalized feature importance of 1\.9% and 1\.4% for VAANI and WAXAL, respectively\.

### 7\.2Phylogenetic relatedness is not sufficient\.

Phylogenetic relatedness is a useful starting point for donor selection but isn’t sufficient to explain transfer performance across either dataset\. In VAANI\-D, several closely related language pairs, such as Awadhi \(awa\) – Chhattisgarhi \(hne\) and Garhwali \(gbm\) – Kumaoni \(kfy\), perform well under both DonorRank and genealogy\-based selection\. However, other cases show that the nearest linguistic relative is not always the most effective donor\. For example, DonorRank selects Rajasthani \(raj\) rather than Hindi \(hin\) for Haryanvi \(bgc\), and Marwari \(mar\) as the best donor for Hindi itself, yielding the best transcription error rates\. This shows that higher\-resource or more widely spoken languages are not necessarily optimal transfer sources\. WAXAL exhibits a similar pattern\. Closely related pairs such as Amharic–Tigrinya are successfully identified, while for other targets, DonorRank selects donors that differ from the nearest genealogical neighbor and yield improved transfer performance\. For instance, Shona, a Bantu language, is the highest\-ranked donor for Malagasy, an Austronesian language; these are completely unrelated languages phylogenetically, yet DonorRank’s selection of Shona results in lower transcription error rates \(e\.g\., CER−10\-10points\)\. These results show that genetic similarity is informative, but transferability is ultimately influenced by multiple linguistic and corpus\-level factors\.

### 7\.3Successful transfer is characterized by complementary signals\.

Features IncludedMean NDCGHrsPhonGenSynGeoOvlpTTRWAXALVAANISingle Features✓✓✓0\.865–✓0\.8700\.952✓0\.9350\.945✓0\.9370\.912✓0\.9530\.950Multi\-Feature Combinations✓✓✓0\.9380\.957✓✓✓0\.9460\.958✓✓✓0\.9510\.960✓✓✓0\.9530\.963✓✓✓✓0\.9470\.962Full Dataset Models✓✓✓✓✓✓✓0\.9480\.986Table 4:Ablation study on best performing DonorRank framework across WAXAL and VAANI\-D datasets\. Hrs=Training and Testing Hours, Phon=Phonological, Gen=Genetic, Syn=Syntactic, Geo=Geographical, Ovlp=Overlap Percentage, TTR=Type\-to\-Token Ratio\.To understand the role of individual features, we perform ablation experiments by training DonorRank using different subsets of linguistic and resource features \(Table[4](https://arxiv.org/html/2608.11441#S7.T4)\)\. No single feature fully explains transfer, although several individual features produce strong rankings\. In VAANI\-D, phonological features provide the best performance \(0\.952\), followed by geographic \(0\.950\) and genetic features \(0\.945\)\. In WAXAL, geographic features are the strongest individual predictor \(0\.953\), while syntactic and genetic features also perform well \(0\.937 and 0\.935\)\. The strongest cue therefore differs across the two collections, while multiple forms of linguistic similarity remain informative in both\. Combining linguistic features improves ranking performance beyond most individual features\. In VAANI\-D, the best linguistic combination reaches 0\.963, compared with 0\.952 for the strongest single feature, while the full model achieves the highest overall NDCG of 0\.986\. In WAXAL, several linguistic combinations closely match the strongest individual geographic model\. The full model remains competitive at 0\.948, but doesn’t outperform the best reduced configurations\. These results show that transferability presents differently for the two corpora\. VAANI\-D benefits from linguistic and resource information, whereas WAXAL can be modeled using a smaller set of linguistic cues\. More broadly, the ablations show that donor selection is not reducible to a single definition of language similarity, and that DonorRank can expose which combinations of signals are useful for a particular multilingual composition\.

### 7\.4DonorRank discovers transfer hubs and widely effective donor languages\.

![Refer to caption](https://arxiv.org/html/2608.11441v1/donor_langs.png)Figure 4:Distribution of languages in DonorRank’s top\-5 donors on WAXAL and VAANI\-D\.We also find that some languages consistently rank highest as donors across many targets\. These languages act as transfer hubs in their respective collections\. In VAANI\-D, languages like Chhattisgarhi \(hne\) and Magadhi \(mag\) emerge as effective donors despite not being the highest\-resource or most mainstream languages\. Similarly, languages such as Shona \(sna\), Maasai \(mas\), and Luganda \(lug\) rank as the strongest donors for a wide range of targets, though they aren’t as widely spoken as other languages in the WAXAL dataset\. This suggest that some languages occupy central positions within the transfer landscape of a multilingual collection\. Identifying such transfer hubs offers practical guidance for expanding multilingual ASR systems to new low\-resource languages, when prior transfer experiments are unavailable\.

## 8Conclusion

We presented DonorRank, a learning\-to\-rank framework for donor language selection in low\-resource cross\-lingual ASR\. Using two complementary datasets, we show that effective donor language selection can be learned in both closely related and typologically diverse language compositions\. Beyond donor selection, DonorRank is a framework for understanding multilingual transfer itself\. Our analyses show that the linguistic cues associated with successful transfer depend on the composition of the multilingual dataset, while highlighting that transferability can’t be explained by any single notion of similarity\. Instead, effective donor selection emerges from multiple complementary linguistic and corpus\-level signals\. As multilingual speech resources continue to expand to under\-studied languages and language communities, we hope DonorRank provides both a practical approach for donor language selection and a useful framework for analyzing transfer in future multilingual ASR collections\.

## Limitations

Our experiments were focused on Indic and African languages, and may not generalize to all language families\. Additionally, our evaluation focuses on a fixed fine\-tuning setup using the w2vBERT model, and other training strategies may affect transfer performance\. Lastly, missing linguistic features for some languages in the lang2vec database reflecting the real\-world gaps in documentation may influence our findings\.

## References

- Blaschkeet al\.\(2025\)V\. Blaschke, M\. Fedzechkina, and M\. ter HoeveAnalyzing the effect of linguistic similarity on cross\-lingual transfer: tasks and input representations matter\.InACL,External Links:[Link](https://arxiv.org/abs/2501.14491)Cited by:[§2](https://arxiv.org/html/2608.11441#S2.SS0.SSS0.Px3.p1.1)\.
- Chenet al\.\(2024\)J\. Chen, C\. Chu, S\. Li, and T\. KawaharaData selection using spoken language identification for low\-resource and zero\-resource speech recognition\.In2024 Asia Pacific Signal and Information Processing Association Annual Summit and Conference \(APSIPA ASC\),Vol\.,pp\. 1–6\.External Links:[Document](https://dx.doi.org/10.1109/APSIPAASC63619.2025.10848811),[Link](https://sap.ist.i.kyoto-u.ac.jp/EN/bib/intl/CHE-APSIPA24.pdf)Cited by:[§2](https://arxiv.org/html/2608.11441#S2.SS0.SSS0.Px2.p1.1)\.
- Chunget al\.\(2021\)Y\. Chung, Y\. Zhang, W\. Han, C\. Chiu, J\. Qin, R\. Pang, and Y\. WuW2v\-BERT: combining contrastive learning and masked language modeling for self\-supervised speech pre\-training\.In2021 IEEE Automatic Speech Recognition and Understanding Workshop \(ASRU\),Vol\.,pp\. 244–250\.External Links:[Document](https://dx.doi.org/10.1109/ASRU51503.2021.9688253)Cited by:[§4\.1](https://arxiv.org/html/2608.11441#S4.SS1.p1.1)\.
- Collins \(2009\)C\. CollinsSyntactic structures of the world’s languages \(sswl\)\.InSyntactic Structures of the World’s Languages Database,External Links:[Link](https://api.semanticscholar.org/CorpusID:61893942)Cited by:[§4\.2](https://arxiv.org/html/2608.11441#S4.SS2.p1.1)\.
- Communicationet al\.\(2023\)S\. Communication, L\. Barrault, Y\. Chung, M\. C\. Meglioli, D\. Dale, N\. Dong, M\. Duppenthaler, P\. Duquenne, B\. Ellis, H\. Elsahar, J\. Haaheim, J\. Hoffman, M\. Hwang, H\. Inaguma, C\. Klaiber, I\. Kulikov, P\. Li, D\. Licht, J\. Maillard, R\. Mavlyutov, A\. Rakotoarison, K\. R\. Sadagopan, A\. Ramakrishnan, T\. Tran, G\. Wenzek, Y\. Yang, E\. Ye, I\. Evtimov, P\. Fernandez, C\. Gao, P\. Hansanti, E\. Kalbassi, A\. Kallet, A\. Kozhevnikov, G\. M\. Gonzalez, R\. S\. Roman, C\. Touret, C\. Wong, C\. Wood, B\. Yu, P\. Andrews, C\. Balioglu, P\. Chen, M\. R\. Costa\-jussà, M\. Elbayad, H\. Gong, F\. Guzmán, K\. Heffernan, S\. Jain, J\. Kao, A\. Lee, X\. Ma, A\. Mourachko, B\. Peloquin, J\. Pino, S\. Popuri, C\. Ropers, S\. Saleem, H\. Schwenk, A\. Sun, P\. Tomasello, C\. Wang, J\. Wang, S\. Wang, and M\. WilliamsonSeamless: multilingual expressive and streaming speech translation\.External Links:2312\.05187,[Link](https://arxiv.org/abs/2312.05187)Cited by:[footnote 1](https://arxiv.org/html/2608.11441#footnote1)\.
- de Vrieset al\.\(2022\)W\. de Vries, M\. Wieling, and M\. NissimMake the best of cross\-lingual transfer: evidence from POS tagging with over 100 languages\.InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),S\. Muresan, P\. Nakov, and A\. Villavicencio \(Eds\.\),Dublin, Ireland,pp\. 7676–7685\.External Links:[Link](https://aclanthology.org/2022.acl-long.529/),[Document](https://dx.doi.org/10.18653/v1/2022.acl-long.529)Cited by:[§2](https://arxiv.org/html/2608.11441#S2.SS0.SSS0.Px3.p1.1)\.
- Diacket al\.\(2026\)A\. Diack, P\. Nelson, K\. Agbesi, A\. Nakalembe, M\. MohamedKhair, V\. Dube, T\. Siyavora, S\. Venugopalan, J\. Hickey, U\. Okonkwo, A\. Bapna, I\. Wiafe, R\. D\. Helegah, E\. D\. Atsakpo, C\. Nutrokpor, F\. B\. P\. Winful, K\. K\. Solaga, J\. Abdulai, A\. O\. Ekpezu, A\. Niyonkuru, S\. Rutunda, B\. Ishimwe, M\. Melese, E\. Bainomugisha, J\. Nakatumba\-Nabende, A\. Katumba, C\. Babirye, J\. Mukiibi, V\. Kimani, S\. Kibacia, J\. Maina, F\. Emmah, A\. I\. Shekarau, I\. S\. Adamu, Y\. Abdullahi, H\. Lakougna, B\. MacDonald, H\. Shemtov, A\. Walcott\-Bryant, M\. Cisse, A\. Hassidim, J\. Dean, and Y\. MatiasWAXAL: a large\-scale multilingual african language speech corpus\.External Links:2602\.02734,[Link](https://arxiv.org/abs/2602.02734)Cited by:[§1](https://arxiv.org/html/2608.11441#S1.p3.1),[§3](https://arxiv.org/html/2608.11441#S3.p1.1)\.
- M\. S\. Dryer and M\. Haspelmath \(Eds\.\) \(2013\)M\. S\. Dryer and M\. Haspelmath \(Eds\.\)WALS online \(v2020\.4\)\.Data set,Zenodo\.External Links:[Link](https://doi.org/10.5281/zenodo.13950591),[Document](https://dx.doi.org/10.5281/zenodo.13950591)Cited by:[§4\.2](https://arxiv.org/html/2608.11441#S4.SS2.p1.1)\.
- Ebrahimiet al\.\(2025\)A\. Ebrahimi, A\. Wiemerslage, and K\. von der WenseModel\-based ranking of source languages for zero\-shot cross\-lingual transfer\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,C\. Christodoulopoulos, T\. Chakraborty, C\. Rose, and V\. Peng \(Eds\.\),Suzhou, China,pp\. 32416–32461\.External Links:[Link](https://aclanthology.org/2025.emnlp-main.1650/),[Document](https://dx.doi.org/10.18653/v1/2025.emnlp-main.1650),ISBN 979\-8\-89176\-332\-6Cited by:[§2](https://arxiv.org/html/2608.11441#S2.SS0.SSS0.Px1.p1.1)\.
- Florianet al\.\(2026\)A\. Florian, C\. J\. Amol, H\. K\. Ombaba, X\. Cui, B\. Mwau, B\. M\. Kamau, L\. D\. A\. Wanzare, C\. Fellbaum, and H\. BuzaabaEvaluating the effect of linguistic relatedness on cross\-lingual transfer in large multilingual automatic speech recognition\.External Links:2607\.04814,[Link](https://arxiv.org/abs/2607.04814)Cited by:[§2](https://arxiv.org/html/2608.11441#S2.SS0.SSS0.Px3.p1.1)\.
- Jain and Bhowmick \(2025\)P\. Jain and A\. BhowmickComparative performance analysis of end\-to\-end ASR models on Indo\-Aryan and Dravidian languages within India’s linguistic landscape\.EURASIP Journal on Audio, Speech, and Music Processing2025,pp\.\.External Links:[Document](https://dx.doi.org/10.1186/s13636-025-00395-5),[Link](https://link.springer.com/article/10.1186/s13636-025-00395-5)Cited by:[§4\.1](https://arxiv.org/html/2608.11441#S4.SS1.p1.1)\.
- Järvelin and Kekäläinen \(2002\)K\. Järvelin and J\. KekäläinenCumulated gain\-based evaluation of ir techniques\.ACM Trans\. Inf\. Syst\.20\(4\),pp\. 422–446\.External Links:ISSN 1046\-8188,[Link](https://doi.org/10.1145/582415.582418),[Document](https://dx.doi.org/10.1145/582415.582418)Cited by:[§5](https://arxiv.org/html/2608.11441#S5.SS0.SSS0.Px2.p1.1)\.
- Keet al\.\(2017\)G\. Ke, Q\. Meng, T\. Finley, T\. Wang, W\. Chen, W\. Ma, Q\. Ye, and T\. LiuLightGBM: a highly efficient gradient boosting decision tree\.Advances in Neural Information Processing Systems30\.Cited by:[§4\.1](https://arxiv.org/html/2608.11441#S4.SS1.p2.1)\.
- Linet al\.\(2019\)Y\. Lin, C\. Chen, J\. Lee, Z\. Li, Y\. Zhang, M\. Xia, S\. Rijhwani, J\. He, Z\. Zhang, X\. Ma, A\. Anastasopoulos, P\. Littell, and G\. NeubigChoosing transfer languages for cross\-lingual learning\.InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics,A\. Korhonen, D\. Traum, and L\. Màrquez \(Eds\.\),Florence, Italy,pp\. 3125–3135\.External Links:[Link](https://aclanthology.org/P19-1301/),[Document](https://dx.doi.org/10.18653/v1/P19-1301)Cited by:[§2](https://arxiv.org/html/2608.11441#S2.SS0.SSS0.Px1.p1.1),[§2](https://arxiv.org/html/2608.11441#S2.SS0.SSS0.Px4.p1.1),[§4\.1](https://arxiv.org/html/2608.11441#S4.SS1.p2.1),[§4](https://arxiv.org/html/2608.11441#S4.p1.1)\.
- Littellet al\.\(2017\)P\. Littell, D\. R\. Mortensen, K\. Lin, K\. Kairis, C\. Turner, and L\. LevinURIEL and lang2vec: representing languages as typological, geographical, and phylogenetic vectors\.InProceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers,M\. Lapata, P\. Blunsom, and A\. Koller \(Eds\.\),Valencia, Spain,pp\. 8–14\.External Links:[Link](https://aclanthology.org/E17-2002/)Cited by:[§2](https://arxiv.org/html/2608.11441#S2.SS0.SSS0.Px1.p1.1),[§4\.2](https://arxiv.org/html/2608.11441#S4.SS2.p1.1)\.
- Mitsumoriet al\.\(2025\)S\. Mitsumori, S\. Kashiwagi, K\. Tanaka, and S\. MorishimaCross\-lingual Data Selection Using Clip\-level Acoustic Similarity for Enhancing Low\-resource Automatic Speech Recognition\.InInterspeech 2025,pp\. 3314–3318\.External Links:[Document](https://dx.doi.org/10.21437/Interspeech.2025-1399),ISSN 2958\-1796,[Link](https://www.isca-archive.org/interspeech_2025/mitsumori25_interspeech.html)Cited by:[§2](https://arxiv.org/html/2608.11441#S2.SS0.SSS0.Px2.p1.1)\.
- S\. Moran and D\. McCloy \(Eds\.\) \(2019\)S\. Moran and D\. McCloy \(Eds\.\)PHOIBLE 2\.0\.Max Planck Institute for the Science of Human History,Jena\.External Links:[Link](https://phoible.org/)Cited by:[§4\.2](https://arxiv.org/html/2608.11441#S4.SS2.p1.1)\.
- Riceet al\.\(2025\)E\. Rice, A\. Marashian, H\. Haynie, K\. von der Wense, and A\. PalmerUntangling the influence of typology, data, and model architecture on ranking transfer languages for cross\-lingual POS tagging\.InProceedings of the 1st Workshop on Language Models for Underserved Communities \(LM4UC 2025\),S\. Truong, R\. A\. Putri, D\. Nguyen, A\. Wang, D\. Ho, A\. Oh, and S\. Koyejo \(Eds\.\),Albuquerque, New Mexico,pp\. 22–31\.External Links:[Link](https://aclanthology.org/2025.lm4uc-1.4/),[Document](https://dx.doi.org/10.18653/v1/2025.lm4uc-1.4),ISBN 979\-8\-89176\-242\-8Cited by:[§2](https://arxiv.org/html/2608.11441#S2.SS0.SSS0.Px1.p1.1)\.
- Sanet al\.\(2024\)N\. San, G\. Paraskevopoulos, A\. Arora, X\. He, P\. Kaur, O\. Adams, and D\. JurafskyPredicting positive transfer for improved low\-resource speech recognition using acoustic pseudo\-tokens\.InProceedings of the 6th Workshop on Research in Computational Linguistic Typology and Multilingual NLP,M\. Hahn, A\. Sorokin, R\. Kumar, A\. Shcherbakov, Y\. Otmakhova, J\. Yang, O\. Serikov, P\. Rani, E\. M\. Ponti, S\. Muradoğlu, R\. Gao, R\. Cotterell, and E\. Vylomova \(Eds\.\),St\. Julian’s, Malta,pp\. 100–112\.External Links:[Link](https://aclanthology.org/2024.sigtyp-1.13/)Cited by:[§2](https://arxiv.org/html/2608.11441#S2.SS0.SSS0.Px2.p1.1)\.
- VAANI \(2025\)VAANIVAANI: capturing the language landscape for an inclusive digital India \(phase 1\)\.Note:[https://vaani\.iisc\.ac\.in/](https://vaani.iisc.ac.in/)Cited by:[§1](https://arxiv.org/html/2608.11441#S1.p3.1),[§3](https://arxiv.org/html/2608.11441#S3.p1.1)\.

## Appendix AAppendix

### A\.1Feature Importance for DonorRank

In Table[5](https://arxiv.org/html/2608.11441#A1.T5), we list the cumulative feature importance for both the CER\-based ranker\. Geographical features of the donor language emerge as the most important factor in determining successful donor transfer\.

Table 5:Feature importance for CER\-based \(best\) ranker for VAANI\.We see syntactic features of both the target and donor language emerge as the most important features for the WER\-based ranker on the WAXAL dataset in Table[6](https://arxiv.org/html/2608.11441#A1.T6)\. The geographical features are the second most important factor determining cross\-lingual transfer\. Dataset features emerge as less important than these linguistic features likely due to

Table 6:Feature importance for WER\-based \(best\) ranker for WAXAL\.The Table[7](https://arxiv.org/html/2608.11441#A1.T7)lists the most important individual features across the different feature groups for the CER\-based DonorRank on the VAANI dataset\. Lexical overlap between the target language and the donor language emerges as the most important factor, followed by the number of training and testing hours\. We find that certain indices in the geographical vectors from lang2vec constantly rank amongst the most important features for determining successful cross\-lingual transfer\.

Table 7:Top 20 features for CER\-based ranker on the VAANI dataset\. HM==headmark, N==noun, Neg\.==negation\.
### A\.2DonorRank Cross Validation results on VAANI\-D and WAXAL

In Table[8](https://arxiv.org/html/2608.11441#A1.T8), we report the number of hours of training data and testing data available for each language\. We then report the top\-5 donors chosen by DonorRank and the NDCG score for each unseen target language held out for cross\-validation\.

Table 8:Train/test hours \(Tr/Te, capped at 7 hours\), DonorRank top\-5 donor languages, and NDCG scores across target languages, for VAANI\-D \(left\) and WAXAL \(right\), using CER as the metric to train the ranker\.