50 Common Words AI (and People) Mishear Across the English-Speaking World

This post explores British and American pronunciation differences through the lens of AI speech recognition, offering new insight into how we speak and how machines listen.

As someone who teaches English phonetics in Singapore, I’m constantly navigating the subtle differences in pronunciation across accents. But nothing has opened my eyes to those differences quite like trying to get AI to understand them. Training language models has become a new lens, a strange and fascinating window, into how speech works, and where it breaks down. And although we take a deep dive into the research in this post, it’s important to appreciate how difficult this challenge really is—I know firsthand as a teacher.

You: “Set a timer for my next shed-jool.”
Siri: “Searching the web for… ‘shuttle?’”

From shed-jools to vitamins, even the smartest AIs struggle to keep up with how English is spoken around the world. The reason? Speech models aren’t trained on all English equally.

Speech AI now powers your Zoom meetings, call centers, captions, classrooms, and health apps. But these tools are overwhelmingly trained on standard American English and it shows. A Stanford-led study found that African-American speech had nearly double the word error rate (35%) compared to white American speech (19%). A separate global benchmark showed that accents from India, Britain, and France had up to 49% more errors compared to American-accented input.

Even OpenAI’s Whisper, one of the best available today, still performs better on U.S. English than on British or Australian. Meanwhile, platforms like Siri and Alexa have long struggled with Irish and Scottish accents, prompting users to change how they speak just to be understood.

This isn’t about funny mishearings, it’s about fairness, inclusion, education, and opportunity. When your accent isn’t recognized, your message often isn’t either.

And the systems aren’t being rude. They’re just undertrained.

Until models are exposed to a wider range of global speech, millions of users will keep getting subtly sidelined by systems that just don’t “hear” them.

Why AI Still Trips Over Accents

Even today’s most advanced speech systems struggle to understand all accents equally well. Despite impressive improvements, multiple studies show consistent accent bias, particularly against non-American and non-standard dialects.

The Data: Bias in the Benchmarks

As mentioned above, a Stanford-led evaluation (Koenecke et al., 2020) of five major ASR systems—Google, Apple, IBM, Microsoft, and Amazon, found that transcripts for African American Vernacular English had a 35% word error rate (WER), nearly double the 19% for white American speakers.

In a larger audit spanning 2,700 speakers across five continents, DiChristofano et al. (2023) found WER gaps ranging from 2–12% for non-American accents, translating to up to 49% relative error. Indian, French, and Southeast Asian-accented English were among the hardest hit.

Even among native speakers, bias persists. OpenAI’s Whisper performs far better on American English than on British or Australian varieties (Graham & Roll, 2024). UK studies show that “prestige” accents like Received Pronunciation are recognized more accurately than regional dialects (Markl, 2022).

Why It Happens

Most speech models are trained predominantly on American data. As a result, the burden of intelligibility often falls on the speaker not the system.

Speech recognition models work by breaking down audio into phonemes and matching them against statistical patterns. But those patterns reflect the training data. And when that data skews heavily toward a single accent, everyone else is left misheard.

Sound Patterns That Commonly Confuse AI

Certain phonetic shifts are especially tricky for ASR systems:

• /æ/ becomes /ɑː/ in words like advertisement, patent, and pasta. Vowel shifts are unevenly represented across training corpora.

• /aɪ/ vs. /ɪ/ in words like either, neither, and vitamin. Models often guess based on spelling alone, lacking contextual accent knowledge.

• Non-rhotic /r/ drops in British English words like parliament, wrath, and version, leading to alignment errors.

• /ʃ/ vs. /sk/ in schedule (shed-yool vs. sked-jool). This cluster confusion breaks decoder predictions.

• /ɪə/ vs. /ɛ/ in leisure and niche. American decoders often miss the glide.

• T-flapping vs. T-holding in tomato and route. American English flaps the /t/; British English pronounces it clearly—confusing the ASR.

These sound mismatches often trip up models in real-world use. For instance, schedule, vitamin, and mobile are common stumbling blocks for AI depending on whether it hears British or American pronunciation.

Anecdotal Proof from the Real World

This isn’t just theory. Users have long reported needing to “code-switch” or Americanize their speech just to be understood:

• Siri and Alexa had well-documented trouble with Scottish and Irish accents.

• Google Assistant often mishears British time phrases due to differences in /t/ pronunciation.

• Contact center pilots have shown customer satisfaction increases when using accent-neutralizing AI layers.

• Reddit users frequently report toggling between British and American spellings or pronunciations to get accurate results.

Until these solutions become standard, speech recognition remains biased. For teachers, students, and users with regional or global accents, the experience can still feel exclusionary. The irony is sharp: the AI that talks back still doesn’t always listen properly.

The Ultimate British vs. American Pronunciation Table

What confuses humans also confuses machines. Below is a table of over 50 words with distinct British and American pronunciations, many of which AI often misunderstands, especially when the user’s accent doesn’t match its assumptions.

The Real-World Impact of British and American Pronunciation Differences

Why does it matter if AI misunderstands your pronunciation?

Because mistakes aren’t just funny. They affect money, inclusion, and opportunity:

Sales Calls:

Real-time accent-conversion technology is already reshaping customer interactions. Krisp’s Accent Conversion (2025) and Sanas have shown that simply adjusting a rep’s accent (without changing their voice identity) can improve comprehension and comfort — resulting in a 15% to 26% boost in sales conversions, depending on the pilot. These tools let people retain their personality while becoming clearer to global customers.

Education Apps:

In language-learning contexts, many AI tools are unintentionally skewed toward American English norms. This means that learners who pronounce vitamin or schedule the British way are sometimes flagged as incorrect — undermining confidence. Especially in multilingual countries like Singapore or India, children learning British-based systems are unfairly evaluated by models trained on American audio.

Customer Service:

Call center agents with strong regional accents (e.g. Filipino, Nigerian, Scottish) face higher misrecognition rates from AI phone systems. This leads to longer handling times, more repeated clarifications, and lower First Call Resolution (FCR) rates. According to Qualfon and Sanas, accent-related misunderstandings can reduce Customer Satisfaction (CSAT) scores by up to 20% if not addressed — but rise by over 3.5% with proper accent-neutralization AI.

Agent Turnover:

Repeatedly being misheard is demoralizing. Reports from the contact center industry show that accent bias contributes to burnout and attrition rates above 40% in some teams. Accent training programs are costly and slow — AI conversion tools now offer a less intrusive, tech-first solution.

The way you say “route” (root vs. rout) shouldn’t decide whether you get the right help—or the wrong one.

Good News: AI Is Getting Better at Listening

While accent bias remains a problem, 2025 has brought a wave of innovation that’s helping AI catch up to the way real people actually speak.

New methods are making speech recognition models more flexible, responsive, and inclusive:

• Accent-adaptive fine-tuning allows AI to adjust to new accents without needing full retraining. This technique has cut error rates by up to 25% in global English tests.

• Real-time accent conversion tools, like those developed by Krisp and Sanas, can shift pronunciation from one accent to another while keeping the speaker’s unique voice. The result: improved accuracy without sounding robotic.

• Community-sourced speech datasets, such as Mozilla’s Common Voice, are adding diverse, real-world voices into training pipelines. These crowdsourced recordings help models recognize conversational English instead of just textbook examples.

At the same time, leading models are evolving:

• OpenAI’s Whisper and Google’s ModelArk are learning to adapt more gracefully to accents like Singapore English.

• Speechmatics’ Global English model avoids rigid dialect switching by treating English as a fluid spectrum, improving performance across blended and international speech.

• LLM-enhanced speech systems, like Whisper paired with GPT-4 or Gemini, show greater contextual understanding, especially on less common accents.

The combination of larger, more inclusive datasets, smarter model architectures, and real-time speech adjustment is closing the gap. While progress is uneven and slower in some industries, the direction is clear: AI is finally starting to listen better.

Final Thought

Given how rare it is for even a human to effortlessly understand every accent on Earth, it’s actually remarkable that AI is giving it a decent go. Speech recognition systems have come a long way, and many are learning faster than we expect.

Tomato, tomahto.
Schedule, sked-jool.
Privacy, praivacy.

There’s no wrong way to say it, but your AI doesn’t know that yet.

The words in this list aren’t just quirks, they’re evidence. They show how something as subtle as vowel stress or dropped consonants can quietly exclude people from the full benefits of speech-driven tech.

If we want AI that works for everyone, whether you’re saying vitamin or VIT-uh-min, we need to keep teaching these systems to listen more broadly, more fairly, and more humanly.

Because language isn’t just about words. It’s about inclusion.

Deep research here

References

• Koenecke, A., et al. (2020). Racial disparities in automated speech recognition. Proceedings of the National Academy of Sciences (PNAS).
• DiChristofano, D., et al. (2023). Quantifying Performance Disparities Across Dialects in ASR. ICLR Conference Proceedings.
• Graham, S., & Roll, A. (2024). Accent Bias in OpenAI Whisper: A Comparative Evaluation. Journal of the Acoustical Society of America.
• Markl, N. (2022). Prestige and Stigma in Regional UK Accents and their Impact on ASR. UK Linguistic Studies Quarterly.
• Speechmatics. (2022–2024). Real-Time Speech Recognition Benchmarks and Accent Coverage Whitepapers.
• Sanas. (2024). Accent Conversion Performance Report.
• Qualfon. (2024). Accent Neutralization AI in Offshore Support: CSAT & QA Outcomes.
• Reddit and StackExchange. (2023–2024). User-reported experiences with ASR accent errors.
• Whisper Research Paper. (2023). OpenAI.
• Mozilla Common Voice. (2023). Open Dataset Overview.
• Google ModelArk Technical Blog. (2025). Improving Multilingual and Multi-accent ASR.
• Krisp. (2025). Accent Conversion AI Pilot Case Study.

Daniel T Kerson

AI consultant. Writer. Builder. Based in Singapore for 20 years. He runs three projects at the intersection of technology, language, and creativity.

AngMohDan

Language, culture, and ideas

Musical Brick

LEGO music legends

Kerson AI

AI literacy and tools