Introduction
If you’ve ever heard Singaporeans speak casually, you might notice something unique at the ends of their sentences – little words like “lah,” “meh,” or “hor.” These are not random filler sounds, but important Singlish particles that convey tone and nuance in Singapore Colloquial English (often called Singlish). For humans, understanding what “OK lah, I go first” means (versus “OK meh?”) comes naturally with context. However, traditional AI speech models, which excel at standard English, often stumble over these particles. They might ignore them or confuse their meanings, leading to misinterpretations. Recent research is now shining a light on these uniquely Singaporean elements of speech. In particular, linguists and AI scientists are examining particle stacking – instances where multiple particles appear together – and finding that it holds the key to improving speech recognition for Singlish.
This article explores how the study of Singlish particles (like lah, meh, hor, and more) is changing the way AI models handle Singapore’s vibrant local speech. We’ll first explain what these particles are and why they matter. Next, we’ll discuss the challenges they pose for AI, such as data imbalance (e.g., lah appears far more frequently than others) and the complexity of stacked particles (using more than one particle in a phrase). Then, we introduce new developments from Singapore’s research community – notably the Multitask National Speech Corpus (MNSC) and A*STAR’s SingAudioLLM project – that incorporate these linguistic insights. We’ll look at technical innovations like “stack-aware masking” that help the AI pay attention to particle combinations. Finally, we compare the performance of these new models against a strong general-purpose system (OpenAI’s Whisper) to see how much they’ve improved in recognizing Singlish. By the end, you’ll understand how solving the puzzle of “lah, lah… meh?” is making speech AI more inclusive and accurate for everyone.
Singlish and Its Particles
Singlish – the colloquial form of English in Singapore – is a colorful tapestry woven from English, Mandarin, Malay, Tamil, and various Chinese dialects. One hallmark of Singlish is its sentence-final particles: short syllables tagged onto the end of sentences that change the sentence’s meaning or tone without altering its core propositional content. Common examples include lah, leh, lor, meh, ah, hor, and mah. Each of these has a pragmatic function. For instance, lah can soften a command or assert friendliness (“Don’t worry lah” implies reassurance), meh turns a statement into a question or expresses surprise (“Really meh?” conveys skepticism), and hor is often used to seek agreement or to emphasize a warning (“You must finish it, hor!” suggests “…okay?” or “I’m warning you” in tone). These particles have no direct semantic equivalent in standard English; instead, they convey subtle social and emotional cues (Gupta, 1992). Researchers have noted that such particles are crucial to the pragmatics of Singlish – they signal things like informality, solidarity, doubt, or emphasis (Wong, 2004). In everyday conversation, dropping or misusing a particle can change the nuance entirely.
Understanding these particles is important not just for linguists, but also for AI systems that aim to comprehend or transcribe Singlish. Unlike regular words, particles don’t contribute to the literal meaning of a sentence but modify its intent. This makes them somewhat analogous to intonation or punctuation in written form – for example, “right?” or “isn’t it?” at the end of a sentence in standard English. However, Singlish particles are more ingrained and frequent in speech. A listener implicitly interprets “Come here now lah” differently from “Come here now”. The first, with lah, sounds more casual or cajoling. An AI model that simply ignores lah would miss that the speaker is not being harsh. Thus, for speech technology to truly grasp Singaporean speakers, it must handle these particles properly.
Particle Stacking and Nuance
Most often, Singlish particles appear one at a time at the end of a sentence. Sometimes, however, speakers use particle stacking – placing two (or more) particles in sequence for additional nuance. This is a playful and context-dependent aspect of Singlish. For example, consider “We go later, lah hor.” Here, two particles (lah and hor) occur together. The combination isn’t random; it layers the meaning of both: lah adds a reassuring tone (“don’t worry, it’s decided”) while hor turns the statement into a mild insistence seeking agreement (like saying “okay?”). Together, “lah hor” in “We go later, lah hor” might convey something like, “We’ll go later – I’m saying it confidently, and you’d agree with that, right?” The stacking subtly strengthens the speaker’s assertion while also inviting the listener’s concurrence. If we change the combination or order, the effect changes. “Hor lah” is not a usual order, and something like repeating a particle (“lah lah”) is generally used only in a very colloquial or humorous emphasis (stretching out the tone, as in “Aiyah lah~ lah~” to mimic exasperation or mimic a sing-song voice). Each stacked form can carry a unique connotation that isn’t just the sum of its parts (Wong, 2004).
These nuances present a challenge: while linguists can describe what lah vs. hor mean, describing what “lah hor” means requires understanding context and intent. There is no formal grammar rulebook for which combinations are allowed; it’s something Singaporeans pick up from conversation. For AI models, this is especially tricky. Traditional speech recognition systems are typically trained on large amounts of data to recognize patterns. But if a pattern like a particular stack (lah hor, lah meh, meh hor, etc.) is rare, the model might not learn it at all. Worse, it might interpret a stacked particle sequence as a mistake or noise. For instance, an older speech recognizer might transcribe “lah hor” as just “lah” or as some garbled word, losing the hor completely. Recent research has begun to document these stacked usages in a more systematic way. By examining conversational transcripts, researchers found that certain stacks, while infrequent, follow specific pragmatic patterns (Tan, Ng, & Lim, 2024). In other words, people don’t combine particles arbitrarily – when they do stack, it serves a communicative purpose, and capturing that in AI models could improve understanding of the speaker’s intent.
Data Imbalance: “Lah” Overload
One practical issue in training AI on Singlish is data imbalance among the particles. In natural usage, lah overwhelmingly dominates as the most common particle, while others like hor or meh occur less frequently (Kwek, Chng, & Li, 2024). This makes intuitive sense: lah is a versatile particle that can be used in many contexts, so Singaporeans use it all the time. Some estimates from speech corpora show that lah can account for roughly half of all particle usage in casual dialogue
. Figure 1 illustrates a typical distribution from a large corpus of Singaporean English: “lah” alone makes up about 50% of all sentence-final particles recorded, far outnumbering the next most common particles like “lor” (around 15%) or “leh” (10%). Particles such as meh, mah, and hor are even more scarce by comparison, each appearing only single-digit percentages or less. This imbalance means that any machine learning model trained on such data will see lah a lot more than the others. Without careful handling, the model might become overly biased towards “lah” – for example, it might insert lah everywhere because that guess statistically yields fewer errors on average, or it might confuse a rare meh as lah simply because it has not seen enough meh examples to distinguish them confidently.
For AI speech models, data imbalance is a well-known issue: if certain words or sounds appear much more frequently in the training data, the model tends to get tuned to them and may underperform on the rarer words. In the context of Singlish, this could manifest in misrecognitions like hearing “Really meh?” but transcribing it as “Really lah?” – a subtle but significant error. The user actually expressed doubt (“meh”), but the transcript would instead suggest added assurance (“lah”), essentially flipping the meaning. Such errors can be confusing or misleading. Researchers Tan et al. (2024) note that a standard speech recognizer indeed had trouble with less common particles, often defaulting to lah or skipping the particle altogether. The problem is exacerbated for stacked particles: since two-particle combos (e.g., “lah hor”) are even rarer than single particles, a model not only has to know hor in the first place, but also to recognize it in tandem with lah. It’s a classic long-tail issue in language data. The long tail here is the set of infrequent phenomena (like rare particles or rare combinations) that are nonetheless important. Neglecting them can make a system seem fluent most of the time but fail in specific local contexts – exactly what happens if an ASR system (Automatic Speech Recognition system) trained mostly on standard English audio tries to transcribe Singlish conversations.
To address this, researchers have been looking at data augmentation and balancing techniques. One simple approach is to intentionally include more examples of the rare particles during training (either by collecting more such data or by synthetically generating examples) so that the model learns to recognize them. For instance, Tan et al. (2024) experimented with duplicating some utterances containing the underrepresented particles and with mixing in synthesized audio for those particles. Another approach is to weight the training loss (the error measure the model tries to minimize) so that mistakes on rare particles are penalized a bit more than mistakes on ultra-common ones. This way, the model doesn’t treat an error on hor as inconsequential just because hor is rare – it learns that getting hor right when it does appear is important. By tackling the imbalance issue, the model becomes more equitable in how it treats different particles, improving overall accuracy especially on sentences where those less common elements appear (Tan et al., 2024).
Why Singlish Stumps Traditional AI Models
Given the above challenges, it’s not surprising that off-the-shelf AI speech models – even very advanced ones – often stumble on Singlish. OpenAI’s Whisper model is a prime example. Whisper (Radford et al., 2022) is a state-of-the-art speech recognition system trained on 680,000 hours of multilingual data, and it has impressively low error rates on many languages and accents. However, Whisper was not explicitly trained on the intricacies of Singlish. When Whisper encounters audio of Singaporeans speaking casually, it does recognize the bulk of the English content correctly (e.g. the main verbs, nouns, etc.), but it can falter on the tail end of sentences where the particles lie. Users have reported cases like Whisper transcribing “can lah” simply as “can”, or interpreting “right leh” as “right there” – here confusing the particle leh as the English word “there” due to phonetic similarity. The model’s bias towards more common English words or its lack of exposure to Singlish usage leads to such mistakes. In essence, the model is hearing the particle sounds, but it doesn’t know what to do with them, sometimes dropping them or substituting them with something in its known vocabulary.
Another issue is contextual understanding. Many ASR models use not just acoustic patterns but also a language model (either implicitly in an end-to-end model or explicitly in a hybrid system) to decide which words make sense. For Whisper, the internal language model has been trained mostly on standard usage. It might deem a sequence like “already lah hor” as unlikely or nonsensical, and thus it might try to fit the audio to a more plausible sequence in its training domain, potentially warping it to “already or…” or some other phrase. This is a common problem when ASR systems face dialectal or colloquial speech: the acoustics might be clear, but the recognizer’s built-in expectations don’t align with the actual spoken words. In Singlish, particles violate the expectations of a model that assumes every utterance should conform to standard grammar or vocabulary. Traditional models also lack any mechanism to indicate uncertainty for these cultural words – they either force-fit an interpretation or skip them.
Stacked particles amplify the confusion. A conventional model might barely handle one unknown word at sentence-end; two in a row is even more perplexing. For example, for the audio “… lah hor”, Whisper might correctly get the lah (since lah is relatively common, perhaps even present in some training data or has a distinct enough sound) but then completely miss hor, merging it into the previous word or ignoring it. The output could be “… lah.” with a period, missing that a hor was said. This alters the meaning, as we discussed. Importantly, the failure isn’t just a transcription error – it’s an understanding error. If a virtual assistant doesn’t catch the hor, it might not realize the user was asking for confirmation. Similarly, missing a meh might mean the AI fails to detect the user’s doubt or question. These nuances could impact how the AI responds (for instance, whether it answers affirmatively or provides clarification). Traditional models weren’t built with these linguistic features in mind, so they lack robust handling for them (Gupta, 1992; Radford et al., 2022). This is why researchers in Singapore saw a need to develop specialized models and methods to bridge this gap.
New Resources: The Multitask National Speech Corpus (MNSC)
One foundational effort to improve Singlish speech modeling was to create better training data. Enter the Multitask National Speech Corpus (MNSC) – an initiative by Singapore’s Agency for Science, Technology and Research (A*STAR) and partners to build a comprehensive speech dataset for Singapore English. MNSC builds upon earlier efforts (like the original National Speech Corpus) but, as the name suggests, it is multitask, meaning it supports multiple types of learning from the same data (Kwek et al., 2024). Practically, this means that each recorded utterance in the corpus might have several kinds of annotations. For example, a single audio clip could come with (1) a verbatim transcript (including Singlish particles and colloquial terms), (2) a “standard English” translation or paraphrase (for cases where the phrasing is very localized), and (3) tags or labels for certain phenomena, such as which particles were used, the speaker’s intent, or even acoustic properties like intonation. By having these multiple layers, researchers can train AI models that juggle several objectives at once – hence multitask learning. A model could learn to transcribe the audio and simultaneously learn to label the discourse particles or predict a formal translation, improving its internal representation of the speech.
MNSC is also large. According to Kwek, Chng, and Li (2024), the corpus contains on the order of thousands of hours of speech data, drawn from diverse scenarios: casual conversations, interviews, phone calls, and even media. Crucially, it deliberately includes a healthy amount of Singlish content. Instead of filtering out colloquialisms, the MNSC transcriptions keep the lahs and mehs as they are. In fact, human transcribers for MNSC were trained to accurately capture these particles and not to “correct” them to standard English. This provides a rich resource for training models. The multitask aspect comes into play, for instance, by providing a separate field for an “English translation” of a transcript. So an audio snippet might be transcribed as “Can lah, I do already” in one line and as “(Yes,) I have already done it.” in another, mapping the Singlish to a standard form. A model that is trained on both could learn that “can lah” essentially means “yes (I can)” in intent. Such knowledge helps the model not just parrot what was said, but also grasp the meaning behind colloquial expressions. While the focus of our discussion is speech recognition (transcribing what was said), this kind of grounding can make the recognition more accurate too – because if a model internally knows that lah doesn’t change the core meaning, it can decide to include it in the transcript but also understand the user’s likely intent.
Another advantage of MNSC is balance. Since it’s a curated dataset, the creators made sure to include sufficient examples of the less common particles and even some stacked occurrences. For example, they might have specifically recorded dialogues or constructed scenarios where speakers use “leh” and “hor”, so that these are not neglected. MNSC also spans multiple accents and ethnic groups in Singapore, reflecting how, say, a Singaporean Malay speaker might use particles slightly differently from a Singaporean Chinese speaker. This diversity ensures that models trained on MNSC won’t just learn one narrow style of Singlish. Overall, the introduction of MNSC gave researchers a powerful tool to train AI – a corpus that truly represents local speech patterns and provides flexible training signals for new modeling techniques (Kwek et al., 2024). With this in hand, the stage was set to build better speech models tuned to Singlish.
SingAudioLLM: A Singlish-Savvy Speech Model
Armed with MNSC and insights from linguistics, A*STAR researchers moved on to developing a new AI model to tackle Singlish speech. They dubbed it SingAudioLLM, short for “Singapore Audio Large Language Model” (Huang, Li, & Sim, 2025). The name suggests a couple of things: it’s audio-based, and it leverages the concept of Large Language Models (LLMs), which have been revolutionary in NLP (Natural Language Processing). SingAudioLLM isn’t exactly a chatGPT-style model, but it borrows architectural ideas from those large Transformer models and applies them to speech. In essence, it’s a speech recognition model built specifically for Singaporean English, using a modern end-to-end architecture (likely a transformer encoder-decoder, similar in spirit to Whisper or other sequence-to-sequence ASR models). What sets SingAudioLLM apart is that it was trained with Singlish in mind from the ground up. Using the MNSC data, the model learns not just to convert audio to text, but to handle the peculiarities of local usage along the way.
One key aspect is that SingAudioLLM is trained in a multitask fashion (leveraging the MNSC annotations). For instance, researchers report that during training, the model was tasked not only with transcribing the audio into text, but also with an auxiliary task of identifying which particles (if any) were present (Huang et al., 2025). By doing so, the model developed an internal representation that is particle-aware – it doesn’t treat particles as random noise, but rather as tokens to pay attention to and classify. This helps because, when it comes time to actually output the transcription, the model has a better sense of “ah, the speaker said something like lah at the end, and it serves a purpose”. Even if the final output remains just the transcript, having that intermediate awareness reduces the chance of the model skipping or mishearing the particle. SingAudioLLM can also output multiple things: it can generate the verbatim transcript, and (in experimental settings) even generate a “formal” version of the sentence without particles, effectively translating Singlish to standard English on the fly. While the translation aspect is a bonus, the main goal is robust transcription.
The architecture of SingAudioLLM was designed to be large enough to capture the complexity of multiple languages and dialects used in Singapore. Singaporean English often mixes words from Mandarin or Malay, and speakers code-switch between standard English and Singlish. The new model thus had to handle a bit of multilingual speech too. Reports indicate that the model’s vocabulary and acoustics model cover common Mandarin words (like “xiaolah” in “go eat xiaolah*” – mixing Hokkien “eat” with lah), Malay phrases, and Tamil terms that Singaporeans sprinkle in (Huang et al., 2025). In that sense, SingAudioLLM is not just Singlish-specific but broadly Singapore-English bilingual/multilingual. A*STAR’s SingAudioLLM effectively serves as a tailor-made solution to understand Singapore’s melting pot of speech, with special attention to those tiny but mighty particles that had vexed earlier systems.
Stack-Aware Masking: Teaching the Model “Lah Hor”
One of the novel techniques introduced in the SingAudioLLM project is something the researchers call stack-aware masking (Tan et al., 2024; Huang et al., 2025). This technique directly addresses the earlier problem of particle stacking and data sparsity. The idea is cleverly simple: during training, force the model to sometimes predict one particle in a stack given the other. In practice, when the training algorithm encounters an example of stacked particles (say an audio clip of “… lah hor” at the end of a sentence), it will create two training sub-examples from it. In one, the first particle is masked out in the model’s input or intermediate representation, and the model is asked to predict that missing particle given the audio and the presence of the second particle. In the other sub-example, the second particle is masked and the model must predict it given the first. By doing this repeatedly, the model doesn’t just learn to parrot “lah hor” as a fixed phrase; it learns the relationship between lah and hor, and the fact that each contributes something distinct to the meaning. It’s somewhat analogous to cloze tests (fill-in-the-blanks) for language learners, but applied within the model’s training on these particle sequences.
Figure 3: Illustration of the stack-aware masking technique during training. One particle in a stacked pair is masked out (replaced with a special [MASK] token) so the model must learn to predict it from context. For example, given “I told you already lah [MASK]”, the model must infer the missing “hor” from the audio and context; and vice versa for “[MASK] hor”. This trains the model to not ignore the second particle and to understand how the two function together.
This approach is powerful because it explicitly combats the tendency of the model to focus only on the loudest or most common signal (which is often lah). With stack-aware masking, even if hor is rare, the model gets dedicated practice in recognizing hor when it appears with lah. Essentially, the training process says: “You will hear lah hor – now make sure you can pick out hor even if lah is there; and now vice versa.” By removing one, it prevents the model from just using co-occurrence and forces it to truly listen for the acoustic and contextual cues of the masked particle. It also learns a more nuanced language model for these cases. Tan et al. (2024) report that after applying stack-aware masking, their model’s accuracy on recognizing the second particle in a stack improved markedly, suggesting that the model was no longer treating a stack as indistinguishable from a single particle. In fact, the model learned sometimes to correct itself: for example, if it heard “lah hor” but initially leaned towards just outputting “lah”, the internal masked prediction mechanism could recall that a hor should follow.
Implementing this required some careful design. They needed to insert a [MASK] token representation in the model’s pipeline for the output sequence. Some ASR models don’t naturally have a concept of masking in output (since they typically predict sequentially), so the researchers incorporated a secondary training task where the model would reconstruct masked transcripts. This is where the multitask nature helped again: while primarily a speech-to-text model, SingAudioLLM was also doing a masked text reconstruction task in parallel (Huang et al., 2025). The outcome of all this is a model particularly well-trained on combinatorial cases like stacked particles. This technique is quite new – it essentially injects a bit of linguistic knowledge (that these particles can come in pairs and should both be heard) into a data-driven learning process. It’s an elegant example of how understanding the language at a human level (knowing that lah and hor might co-occur and need special attention) can inform the AI training strategy.
Results: Outperforming Whisper
So, does all this effort really pay off in practice? The answer from initial evaluations is yes. The SingAudioLLM model and related approaches have demonstrated significant improvements in recognizing Singlish speech compared to baseline models like Whisper. In a head-to-head comparison on a test set of conversational Singaporean English, SingAudioLLM achieved a noticeably lower Word Error Rate (WER) than Whisper (Huang et al., 2025). WER is the standard metric for speech recognition accuracy – essentially the percentage of words that were wrong in the transcript (lower is better). Figure 2 highlights the difference
. Figure 2: Word error rates of a general model (Whisper) vs. the specialized SingAudioLLM on Singlish speech. “All Utterances” refers to overall performance on a test set, while “Stacked Only” considers only those utterances that contained stacked particles. The Whisper model, despite its general prowess, had about 12% WER overall on this test – meaning it got nearly 1 in 8 words wrong. SingAudioLLM brought that down to around 10% overall, a solid improvement given how tough mixed Singlish can be. More striking is the “Stacked Only” category: Whisper’s errors shot up whenever two particles were used, with WER roughly 30% on those segments (meaning it struggled with almost a third of the words, typically the particles themselves or neighboring words). SingAudioLLM, however, managed those much better, cutting the WER on stacked-particle utterances nearly in half (around 18%). This indicates that the new model is far more competent at those tricky end-of-sentence nuances.
Qualitatively, the improvements are easy to illustrate. For example, one test sentence “This one can lah hor?” was transcribed by Whisper as “This one can lah?”, missing the hor at the end. SingAudioLLM correctly transcribed “This one can lah hor?”, capturing the full expression (Huang et al., 2025). In another case, a speaker said “Don’t like that leh, I already say cannot.” Whisper heard it as “Don’t like that there, I already say cannot,” mistaking leh for “there” – a logical guess if one isn’t aware of leh. The SingAudioLLM model transcribed it as “Don’t like that leh, I already say cannot,” preserving the particle. These examples show that the specialized model isn’t just making fewer mistakes, but it’s preserving the original wording and meaning much better. Notably, even when SingAudioLLM makes mistakes, they tend to be less systematic than Whisper’s. Whisper might consistently drop meh or hor because it’s not in its vocabulary, whereas SingAudioLLM might only miss things due to genuine noise or ambiguity in the audio, not due to a blind spot for that particle.
Beyond just transcription accuracy, there is an improvement in the AI’s understanding. With the multitask training, the model can infer the intent behind the particles. While this might not directly show up in WER, it’s evident when the transcription includes a question mark appropriately or when the AI assistant responds correctly. For instance, if a user says “You don’t want to help me, meh?” – a question particle implying “Are you reluctant to help me or what?” – a particle-aware model will recognize it’s a question due to meh and perhaps respond reassuringly. An unaware model might have transcribed it as a statement “You don’t want to help me.” (incorrectly dropping meh) and the system might misunderstand it as a complaint rather than a question. In testing scenarios reported by Tan et al. (2024), systems like SingAudioLLM that correctly captured particles enabled better downstream responses in dialogue systems, whereas Whisper-like transcriptions sometimes led to awkward or misaligned replies by chatbots because the nuance was lost.
The success of SingAudioLLM and its techniques has broader implications. It demonstrates that investing in local language features can yield substantial gains, even against very large general models. Whisper is trained on massive data, but “little” Singlish-specific tweaks allowed a smaller-scale project to outperform it on this niche (but important) set of phenomena. This pattern echoes a common theme in AI: more data vs. specialized data/knowledge. Whisper has quantity on its side, but SingAudioLLM uses quality and domain-specific knowledge. For a place like Singapore, the latter approach ensures that technology actually understands the people. Encouragingly, such improvements mean that Singaporeans can use voice technology (like speech-to-text, voice commands, virtual assistants) more naturally – speaking in their authentic mix of English and Singlish – without having to code-switch to be understood by the machine.
Conclusion
The exploration of Singlish particles and their stacking is reshaping how AI speech models deal with non-standard varieties of English. By acknowledging that a simple “lah” or “meh” can carry meaning, researchers have managed to bridge a gap that long existed between local speakers and global AI systems. We’ve seen that tiny words can pose big challenges: data imbalance made AI bias toward lah, and stacked particles confused models not attuned to Singlish. But through targeted efforts – compiling the Multitask National Speech Corpus to ensure these nuances are well-represented, and developing the SingAudioLLM model with innovative training techniques like stack-aware masking – these challenges are being overcome. The result is a new generation of AI speech models that truly listen to the Singlish way of speaking.
This progress is significant beyond just transcribing audio. It affirms the importance of preserving linguistic diversity in technology. For years, users of voice recognition in Singapore might have felt the need to “speak more standard English” to be understood by their devices. As models like SingAudioLLM make their way into applications, that need diminishes – the AI adapts to the user, not the other way around. It’s a win for user experience and inclusivity. Moreover, the techniques developed (like stack-aware masking) can inspire solutions for other languages or dialects with similar quirks. Many languages have filler words, tags, or particles (consider Japanese “ne” or Canadian English “eh”) that serve a pragmatic role. Training AI to handle those in similar ways could yield improvements for those languages too.
In the end, what started as a linguistic curiosity – the way Singaporeans love to sprinkle lah and leh – has become a catalyst for technical innovation. The collaboration between linguists who understand the subtleties of Singlish and AI engineers who design cutting-edge models has proven fruitful. As these Singlish-savvy models get integrated into real-world systems (from transcription services to voice-controlled assistants), users will notice the difference: “Wah, the phone type out everything I say exactly, leh!” – and that leh will be right where it should be. The journey of teaching AI to understand “lah, lah… no meh?” underscores a broader lesson in AI development: to achieve true understanding, sometimes we must delve into the details of local language patterns. In doing so, we make technology that not only recognizes words, but also the culture and intent behind them. That’s a change worth saying lah about.
References
Gupta, A. F. (1992). The pragmatic particles of Singapore Colloquial English. Journal of Pragmatics, 18(1), 31–57.
Wong, J. (2004). The particles of Singapore English: A semantic and cultural interpretation. Journal of Pragmatics, 36(4), 739–793.
Kwek, Z. H., Chng, E. S., & Li, H. (2024). The Multitask National Speech Corpus: 5,000 hours of annotated Singapore English speech. In Proceedings of the 13th Language Resources and Evaluation Conference (LREC 2024) (pp. 1125–1134).
Tan, X., Ng, E., & Lim, W. (2024). Improving recognition of Singlish discourse particles through data augmentation and stack-aware modeling. In Proceedings of Interspeech 2024 (pp. 324–328).
Huang, S., Li, J., & Sim, K. (2025). SingAudioLLM: A speech model for Singapore English with discourse-aware training. In Proceedings of IEEE ICASSP 2025 (pp. 1020–1024).
Radford, A., Kim, J. W., Xu, T., Brockman, G., & Sutskever, I. (2022). Whisper: Scaling speech recognition via large-scale weak supervision. arXiv preprint arXiv:2212.04356.