Speech deepfake detectors can perform extremely well when they are trained and tested on one benchmark, yet perform poorly on recordings from other sources. This raises a concern: a detector may learn differences between datasets rather than a general distinction between genuine and deepfake speech. This thesis tests whether three detectors—RawNet2, LCNN, andWav2Vec-AASIST—use such dataset-related shortcuts.
First, a data audit of four public datasets shows that simple classifiers can identify which dataset a recording came from using several kinds of audio measurements. This establishes that source-related clues are available, but not that the deepfake detectors use them. To test detector reliance directly, the thesis introduces the mixing ratio protocol. The protocol changes which of two datasets supplies the genuine and deepfake recordings during training, while the evaluation recordings remain fixed. If a detector uses dataset source as a shortcut, reversing this relationship should cause its decisions to change.
This is what occurs. Across all three detectors, reversing the training composition moves the equal error rate (EER) on the same test recordings from 0.003–0.028 to 0.78–0.87. An EER of 0 is perfect, 0.5 is chance, and a value above 0.5 means that the predictions are systematically inverted. A similar response appears with a second dataset pair. By contrast, the results are much flatter when the same procedure is applied to two groups of speakers from one dataset. This control shows that the mixing procedure does not create the main response by itself.
Further experiments test whether one audio property explains the dependence. Reducing the frequency range, matching the average signal level, and removing low-frequency energy do not remove it. Dataset source can also be predicted from several different kinds of measurements. These findings are consistent with source information being represented through several overlapping or untested properties, but they do not identify which properties each detector used. Overall, dataset identity is available as an exploitable shortcut in speech deepfake detection. Future benchmarks should therefore include diverse recording and generation pipelines in both genuine and deepfake classes. ... Read more

Automatic Speech Recognition (ASR) systems perform worse on the speech of deaf and hard-of-hearing (d/Dhh) individuals than on standard speech, creating a substantial accessibility barrier. Improving ASR for this demographic is typically made complicated by severe data scarcity and high inter-speaker heterogeneity. This thesis investigates strategies to reduce the Word Error Rate (WER) for Dutch d/Dhh speech without relying on large, specialized training datasets. First, a prosodic analysis contrasting d/Dhh speakers with standard-hearing controls identifies key deviations in speech duration, loudness, and pitch. To mitigate data scarcity, standard Dutch read speech is augmented to mirror these identified prosodic differences, followed by parameter-efficient fine-tuning (LoRA) on the modified data. While this targeted prosodic augmentation successfully closes the acoustic gap between standard and d/Dhh speech, it does not improve WER, indicating that other differences are the primary drivers of recognition errors. Conversely, transitioning from population-level adaptation to speaker-level personalization yields substantial improvements. Fine-tuning a base model on a small sample of an individual speaker's voice reduces WER, with the most gains observed for the more difficult-to-transcribe speakers. ... Read more

Automatic speech recognition (ASR) remains less accurate for children than for adults. This difference is poorly understood in multilingual systems. This thesis examines whether adding sparse Mixture-of-Experts (MoE) layers to a shared FastConformer-CTC encoder can narrow the adult-child word error rate (WER) gap. It also tests whether experts specialise by age, even without age labels in the training data. Dense and MoE encoders were initialised from the same pretrained checkpoint and fine-tuned in two stages: first on adult speech, then on mixed adult-child speech. Experiments covered English, Dutch, German, and Polish. Paired speaker-level bootstrap tests compared WER gaps. In the English-only experiment, MoE reduced the gap from 22.33 to 19.97 percentage points after adult-only fine-tuning and from 11.05 to 10.69-10.78 points after child-speech adaptation. It also lowered adult and child WER. Part of this advantage emerged before any child speech audio was used, suggesting an architectural contribution beyond data exposure. In the multilingual experiment, the benefit did not generalise uniformly. MoE reduced the gap significantly only for English. Dutch showed a smaller, non-significant reduction. German showed an increase in child WER and a significant widening of the adult-child gap, despite improved adult WER. The Polish results were difficult to interpret, as the corpus contained only isolated words and children with sigmatism. A routing analysis of the English models found age-correlated differences in early and late encoder regions after child-speech fine-tuning. The experts that adult and child speech were routed to produced distinct representations. An English domain-control analysis did not reproduce this pattern. Language was the dominant routing factor, and age-related specialisation was layered on top. Overall, a single shared MoE encoder can reduce the adult-child gap and develop age-related specialisation without demographic labels at inference time. Its effect depends on the coverage and domain match of the child speech data in each language. This makes the sparse MoE encoder a practical direction for more inclusive speech recognition: one model that reduces the adult-child gap without age labels, and whose routing can be examined to explain the improvement. ... Read more

Automatic speech recognition (ASR) systems achieve very low word error rates on standard adult speech but underperform on populations whose speech deviates acoustically from the training distribution. Dysarthric speech, in which neurological injury impairs the muscular control of speech production, is among the most challenging cases: word error rates can exceed 90% on severely impaired speakers even after in-domain fine-tuning. Mixture-of-Experts (MoE) architectures, in which feed-forward sublayers are replaced by parallel “expert” sublayers and a small gating network routes each input to a subset of them, offer a candidate mechanism for inclusive ASR: in principle, the experts could come to specialize on different speaker populations and the router could allocate them where they are needed.

This thesis investigates whether MoE delivers that promise on dysarthric ASR, using the FastConformer-CTC Large encoder fine-tuned per fold on the TORGO and UASpeech corpora across 41 leave-one-speaker-out folds. Three research questions are addressed: how standard sparsely-gated MoE compares to a fine-tuned dense baseline (RQ1); what specialization the trained experts develop (RQ2); and whether either training-time supervision or inference-time router steering can push the model toward useful specialization (RQ3).

The answer to all three questions is negative. Five standard MoE configurations match but do not beat the fine-tuned dense baseline, and a parameter-equal follow-up shows that at equal feed-forward capacity, the MoE structure is actively worse than dense by up to 2.89 pp. The trained experts develop genuine functional specialization, but the axis of specialization is encoder depth, not speaker group or phonetic content: different layers prefer different expert pairs, and within any layer, routing is roughly content-independent and group-independent. Training-time supervision toward group-aware specialization collapses at inference, as the depth-based organization overrides the supervised partition. Inference-time router steering produces real per-speaker improvements under post-hoc selection of the bias strength, but neither of two deployment-realistic mechanisms for choosing that bias recovers the gain.

The MoE structure on this task does meaningful work, but not the work the task needs: it organizes the encoder’s computation along an axis the dense baseline already covers through its per-layer feed-forward weights. Routing-input variants, cross-architecture replication, and continual-learning extensions are flagged as the most direct follow-ups. ... Read more

While Automatic Speech Recognition (ASR) systems have achieved near-human performance for typical adult speech, they remain highly unreliable for children with atypical language development, such as Developmental Language Disorder (DLD). A major bottleneck in improving ASR for this demographic is the severe scarcity of annotated speech data due to strict privacy and ethical constraints. To address this, this study investigates the use of Text-to-Speech (TTS) synthetic data augmentation to improve ASR performance for Dutch children with DLD. We propose a linguistically motivated pipeline that applies rule-based morphological and phonological transformations to typical Dutch child transcripts, simulating realistic DLD error patterns. These transformed transcripts are then synthesized using a zero-shot multi-speaker TTS model (XTTS v2) to generate a controlled dataset of DLD-like child speech. By augmenting ASR training data under various experimental conditions, we evaluate the impact of synthetic phoneme error probability and speaker variability on downstream recognition accuracy.

Our results demonstrate that applying a moderate probability of linguistically motivated DLD transformations significantly outperforms standard, error-free childlike synthetic speech. Despite this relative linguistic success, the synthetic fine-tuning did not surpass the absolute performance of the unadapted zero-shot baseline. Furthermore, experiments indicate that scaling synthetic speaker variability introduces disruptive generation artifacts rather than improving model robustness. These findings demonstrate the validity of targeted linguistic simulation for atypical speech while exposing the severe acoustic limitations of current zero-shot TTS models in clinical, data-scarce domains. ... Read more

Modern Automatic Speech Recognition (ASR) models show reduced performance on speech produced by children with Developmental Language Disorder (DLD) due to data scarcity, pronunciation variability and the different linguistic characteristics. This study investigates whether data augmentation, specifically converting Dutch dysarthric adult speech into Dutch child speech via monolingual Voice Conversion (VC), can improve ASR performance for Dutch children with DLD. This is motivated by overlapping acoustic characteristics between dysarthria and DLD. MeanVC, a zero-shot model was used to convert dysarthric adult speech into child speech. Subsequently Whisper-V3-Large was used to fine-tune on various configurations of synthetic VC speech and real Dutch DLD child speech. While models fine-tuned on VC speech outperformed the zero-shot baseline when real DLD data was unavailable, it was revealed that unconverted dysarthric speech outperformed the VC-generated speech, suggesting that the VC process might have difficulties keeping the necessarily speech characteristics. The study concludes that while dysarthric speech has relevant acoustic characteristics, this specific VC pipeline does not meaningfully improve ASR performance for DLD. ... Read more

Automatic Speech Recognition (ASR) systems are increasingly used in real-world applications, but their performance is not equally reliable across speaker groups. Previous work has shown state-of-the-art ASR systems to perform worse for non-native speakers; the linguistic factors that contribute to this, however, are yet to be understood. To address this gap, this study investigates how utterance length and word category affect ASR performance for Dutch speech from non-native adult speakers. In addition, it investigates how these effects differ between read speech and human-machine interaction (HMI) speech. The study uses ASR outputs from Google Chirp 2 and Whisper large-v3 on non-native adult speech from the JASMIN corpus. The analysis examines how error patterns relate to utterance length and part-of-speech category, with insertion, deletion, and substitution errors analyzed separately for each factor. The results show that ASR performance varies systematically with both factors: very short utterances are especially error-prone, length effects differ between read and HMI speech, and word-category patterns are most visible in substitution and insertion errors. These findings highlight the need for more fine-grained ASR evaluation methods that account for how linguistic context influences recognition errors in non-native speech. ... Read more

Automatic speech recognition systems can be less accurate for children than for older
speakers. For Dutch, previous evaluations found higher word error rates for native children than for native teenagers, but they did not show which types of words contribute to the difference. This paper studies whether that difference is concentrated in particular word categories. A word category is the part of speech assigned to a word in the reference transcript, for example noun, verb, pronoun, article, or preposition. The analysis uses native Dutch child and teenager speech from the JASMIN-CGN corpus in two speaking styles: read speech and speech from human machine interaction tasks. The analysis compares Google Chirp 2 and Whisper-large-v3 outputs with the reference transcripts and groups substitutions and deletions by word category. Insertions are reported separately because an inserted system word has no matching reference word category. Children have higher word error rates than teenagers in every condition. The difference is about 9 percentage points in read speech and about 4 to 5 percentage points in human machine interaction speech. In read speech, most of the substitution and deletion difference comes from nouns, verbs, and pronouns. In human machine interaction speech, errors are spread across categories, so no single category explains much of the difference. A control that compares utterances with the same counts of word categories reduces the verb and pronoun differences, while the noun difference remains largest. A check of the specific words behind these categories shows why this matters: many noun and verb errors come from words found only in the child read reference transcripts, whereas pronoun errors mostly involve words used by both groups. The results show that word categories are useful for locating where the error rate difference appears, but they are not a complete explanation without checking the specific words and speaking material behind
each category. ... Read more

Automatic speech recognition systems achieve near-human performance under standard conditions but perform poorly on dysarthric speech due to high acoustic variability resulting from neuromotor impairment. While speaker-specific adaptation can improve performance, limited training data restricts conventional learning approaches. Contrastive learning offers a promising alternative by encouraging more discriminative phoneme representations from limited data, but its effectiveness depends strongly on how negative examples are selected. This thesis investigates whether personalized contrastive learning can improve Dutch dysarthric phoneme recognition.

A Whisper-based encoder-DNN-CTC model is extended with a triplet-loss objective to improve phoneme-level discrimination. Four negative sampling strategies are compared: randomly selected, phonologically motivated, and two empirically derived from the model's own prediction errors, one estimated on the training set and one via cross-validation. Each is evaluated under two training regimes: contrastive fine-tuning of a pretrained model and training from scratch.

All contrastive approaches significantly outperform a CTC-only baseline. The strongest results are obtained with phonologically motivated and cross-validation-based empirical negatives when training from scratch, yielding up to a 10.7% relative reduction in phoneme error rate. Under fine-tuning, differences between sampling strategies are negligible. In contrast, when trained from scratch, the phonological and cross-validation-based empirical strategies significantly outperform randomly selected and training-set-based empirical negatives.

These findings suggest that, for this speaker, contrastive learning for dysarthric speech benefits from phonologically informed or empirically derived negative pairs rather than random selection. A practical trade-off emerges between the two strongest strategies: phonologically motivated sampling requires no speaker-specific preprocessing and is immediately applicable to new speakers, but generates a large number of triplets and is computationally expensive at training time. Cross-validation-based empirical sampling requires building a speaker-specific confusion matrix upfront, but produces fewer, more targeted triplets and trains more efficiently. Given comparable performance, the choice between them reduces to whether preprocessing overhead or training-time resources are the limiting constraint. ... Read more

Automatic Speech Recognition (ASR) systems perform poorly on speech of children with Developmental Language Disorder (DLD) due to a scarcity of training data and the high variability in speech of these children. This study investigates whether speech created through cross-lingual child-to-child voice conversion (VC) can improve ASR performance for Dutch children with DLD. To this end, multiple data augmentation pipelines were developed using the MeanVC zero-shot VC model, generating synthetic child speech with different source-target speaker pairing strategies (most similar, ran- dom, least similar) and varying augmentation sizes. The augmented datasets were used to fine-tune a Whisper-based ASR model. The results show that fine-tuning the ASR on a combination of synthetic speech and source speech improved ASR performance over an out-of-the-box baseline, but did not lead to better performance than fine-tuning it on real speech of children with DLD alone. Interestingly, speaker-dissimilar VC pairings yielded the best ASR performance. Subjective Mean Opinion Score (MOS) evaluations suggest that synthetic speech was perceived as less natural, less intelligible, and less child-like than real speech and that AMOS predictions only partially align with human judgments and ASR performance. These findings suggest that while cross-lingual voice-converted speech can have value as supplementary training data, improvements in speech quality and voice conversion methods are needed before it can serve as an effective replacement for real speech of a child with DLD in ASR training. ... Read more

Automatic speech recognition (ASR) systems still underperform on speech from Dutch children with developmental language disorder (DLD), largely due to the scarcity of training data. This work investigates whether cross-lingual, child-to-child voice conversion from English children with speech sound disorder can be used to generate synthetic Dutch DLD speech for ASR fine-tuning, and what effect source-target speaker similarity has on the result. We use MeanVC to convert Dutch DLD source speech into three synthetic datasets based on high-, low-, and random-similarity pairings with English target speakers. These datasets are evaluated by fine-tuning Whisper-large-v3 on each and measuring word error rate, and by scoring speech naturalness and intelligibility using the automatic MOS predictor MBNet. Fine-tuning on original Dutch DLD data substantially outperformed all synthetic-only configurations, and adding synthetic data to the original did not improve over original-only fine-tuning. High speaker similarity did not produce the best results, while low-similarity pairings performed unexpectedly well when combined with original data. Contrary to the ASR results, all synthetic datasets scored higher on naturalness and intelligibility than the original recordings, suggesting that these measures are not reliable predictors of ASR training value for DLD child speech. Cross-lingual voice conversion from English SSD child speech does not yet provide a viable route to improving ASR for Dutch children with DLD, though targeted filtering of low-quality synthetic utterances remains a promising direction for future work. ... Read more

Automatic Speech Recognizers (ASR) transcribe speech to text and have become very popular. These systems need to accommodate all types of speech, and problems appear for different genders, ages and regional accents. In this paper we aim to discover what are the differences in types of recognition errors made by Dutch listeners and ASR systems in transcribing Flemish Dutch speech. This has been done through an experiment where human listeners and ASRs listened to 40 audio samples coming from speakers from different regions. Afterwards the transcriptions were collected, processed and we calculated the word error rate (WER). ASRs outperformed humans, scored lower WER, with differences appearing based on speaker region. Results show that ASR systems display similar recognition error patterns, a lot of substitutions and deletions, and very few insertions. The main difference stands in the way humans and ASRs deal with confusing audio. Humans tend to use similar words that fit the context whereas ASRs are more prone to inserting unrelated words. ... Read more

Automatic Speech Recognition (ASR) systems are becoming increasingly common in day-to-day life. Yet, child speech remains challenging for ASR systems. This paper gives the first comparison of Dutch human listeners and Dutch ASR systems on Dutch child speech. It tests whether familiarity with child speech improves human transcription performance and sees if the age of the child speaker influences the transcription performance.

A balanced set of 40 utterances were taken from the JASMIN database (speakers aged 7-11), and were transcribed by 20 humans (10 familiar with child speech (parent/caretaker) and 10 unfamiliar). Transcripts were also gathered from two state-of-the-art ASR systems (Google Telephony and a Conformer model). These transcripts were evaluated against reference transcripts using Word Error Rate (WER). Statistical significance was tested.

Results show that overall ASR transcription performance was comparable to human performance, and in some cases slightly, but not significantly, better. Familiar listeners did not outperform unfamiliar listeners. In fact, there was no significant performance difference between the two groups of humans. Within the 7-11 age range, no clear relationship between speaker age and WER was shown, but results were sensitive to sentence difficulty outliers and "speaker effects". ... Read more

Automatic Speech Recognition (ASR) systems have achieved remarkable performance in recent years, however their robustness against diverse speech, such as teenage speech, remains an area of investigation and research. This study evaluates the performance of a state-of-the-art ASR system (Google Telephony) compared to human listeners in transcribing native Dutch teenage speech. Additionally, it investigates whether a listener's social exposure to teenage speech influences their recognition accuracy. A listening experiment was conducted using a dataset of 40 speech samples of Human Machine Interaction (HMI) speech from native Dutch speakers aged 14 to 16, curated from the JASMIN corpus. The audio samples were transcribed by the ASR model and a group of young adult participants (aged 20--24). Performance was evaluated using Word Error Rate (WER), with a specific focus on the impact of normalizing common Dutch contractions and clitics. The results demonstrate that the ASR system outperformed the average human listener, achieving a lower WER compared to both groups of participants, the one with exposure to teenage speech and the one without. However, human participants with regular social exposure to teenagers performed significantly better on average than those without, confirming that familiarity with the demographic improves recognition accuracy. Furthermore, the analysis reveals that orthographic inconsistencies regarding contractions significantly inflate WERs, with normalization reducing the WERs by quite an extent. These findings suggest that while current ASR models are highly robust for this demographic, human domain knowledge remains a relevant factor in understanding teenage speech speech patterns based on the lower WERs of the humans with exposure to teenage speech than those without. ... Read more

Automatic speech recognition (ASR) systems are used in voice assistants and games. Studies have shown that these ASR systems do not work well for all everyone. E.g., speech from regional accents is recognised worse than the native standard accent and speech from older adults are recognised worse than adults.However, no research has been done in comparing humans and ASR systems on Dutch older adults' speech.This paper will fill this gap by finding out the transcription performance of humans and ASR systems (Google Telephony and a conformer model with XLSR-53 features) on Dutch older adults' speech and comparing the performances with each other.This paper also investigated whether regional accents, gender, and age had a significant effect on the performance of humans and ASR systems.The results showed that humans and ASR systems perform similar to each other and that their performance are positively correlated.The results also showed that humans also do not recognise speech equally from older adults as well. These results suggests that there are speech characteristics for which both humans and ASR systems have trouble with. ... Read more

In an increasingly digital world, the ability of systems to adapt to individual users has become essential. Among cognitive variables informing such systems, the human short-term memory plays a crucial role, as it is responsible for perceiving, storing and retrieving data. This study explores the modalities in which various information about memory-related processes has been used in the field of Human-Centered AI (HCAI) by analyzing input methods, objectives, application domains and nature of implemented adaptations. By conducting a systematic literature review, the paper aims to address this gap, while identifying current trends and challenges. The 45 papers included in this study revealed that increasing or decreasing complexity of content or adjusting difficulty levels based on accuracy rates or time periods can improve human performance and support learning in computer-based environments. Such personalization are especially useful in the educational and healthcare domains. These results provide a basis of design guidelines for future research of adaptive mechanisms considering memory-related information. This survey is anticipated to be a starting point for upcoming developments in this field or succeeding reviews that could incorporate a wider range of the memory scope. ... Read more

Dysarthric speech recognition is challenging due to speech variability caused by neurological disorders. This study explores integrating articulatory features with large pre-trained acoustic model features (e.g., WavLM, Whisper) to improve recognition performance. Different fusion strategies, including concatenation and cross-attention mechanisms, are also compared in this work. Experimental results show that articulatory features can enhance WavLM-extracted features, reducing WER for moderate and mild severity level. t-SNE analysis reveal how articulatory features influence feature representations. These findings highlight the potential of multimodal fusion in improving dysarthric ASR systems. ... Read more

Automatic Speech Recognition (ASR) systems perform well for typical adult speech but remain challenged by children’s speech, especially that of children with Developmental Language Disorder (DLD). This study investigates how ASR performance can be enhanced for DLD speech while maintaining accuracy on typical child speech. Two state-of-the-art ASR models, a conformer-based model and Whisper-large-v3, were evaluated using Dutch typical (Jasmin) and atypical (Auris) child speech. The experiments examine data augmentation methods, including speed perturbation and vocal tract length perturbation, and transfer learning through fine-tuning. Results show that both techniques improve DLD speech recognition without degrading typical speech accuracy. The best performance was achieved by combining augmentation and fine-tuning with domain-matched DLD data, reaching 53.2% WER on the Auris test set, while mismatched fine-tuning reduced gains, particularly for Whisper. Overall, the findings demonstrate that integrating data augmentation and fine-tuning offers an effective, balanced approach toward inclusive and robust ASR for children with DLD. ... Read more

Automatic Speech Recognition (ASR) systems are becoming increasingly popular in this day and age. Unfortunately, due to inherent biases within these systems, performance disparities exist among specific demographic groups. Bias metrics can be used to measure this bias. Within ASR they represent a niche area that has not yet been thoroughly explored. The few bias metrics that exist in literature mainly centre around the performance differences between speaker groups. This paper proposes two new bias metrics that focus not only on performance differences, but also take the base performance into account: Weighted Performance Bias (WPB) and Intergroup Weighted Performance Bias (IWPB). Although the lack of ground truth makes the results less easily interpretable, the results show similar trends within the new metrics as those defined in literature: bias is greatest among non-native Dutch speech. ... Read more

Automatic Speech Recognition is a field that has seen a strong increase in developments in recent years. In order to ensure objectivity and reliability in these systems, it is crucial they remain unbiased and treat speakers equally. This paper explores the bias of two state-of-the-art ASR systems in the domain of Dutch and Flemish speech, specifically towards regional dialects. Specifically, it explores Microsoft's Azure AI Speech Services ASR system and Google Chirp. It analyses the performance of these two systems on the JASMIN-CGN language corpus. The results show that speech from West-Dutch regions is recognized correctly significantly more often than other Dutch regions and speech from Brabant is recognized correctly significantly more often than other Flemish regions. ... Read more