Auto-captions for accented English are usable, but the draft is where the work is. The speech recognizer behind every subtitle generator was trained mostly on a few varieties of English, so a speaker from Chennai, Manila, Lagos or Glasgow gets a draft with predictable slips: names, numbers, brand names, words from a second language, and sound-alike pairs. None of that is fixed by picking a different tool with a bigger accuracy number on its home page. It is fixed by reading the draft with a short checklist before you export.
This guide explains why recognizers struggle with accents, gives a checklist that catches most errors in a few minutes, lists the audio habits that cut errors before they happen, and says when you should pick Hinglish instead of English. One honesty note up front: CapsAI's English has been used most on Indian English. For any other accent, the right test is your own clip through the 3 free minutes, then read the draft.
Why do auto-captions slip on accents?
Three causes account for most errors, and they apply to every tool.
Training data. A recognizer learns from recorded speech paired with transcripts. The large public and commercial corpora are heavy on American and British English, so pronunciations outside those varieties are rarer in the data and more often misheard. A 2021 study of a Dutch speech recognizer by researchers at the University of Amsterdam, the Netherlands Cancer Institute and TU Delft, reported by VentureBeat on 1 April 2021, found that the worst-recognized group of native speakers still had a word error rate around 20% better than the best-recognized group of non-native speakers. The same article cites a Washington Post test in which smart speakers were 30% less likely to understand non-American accents, and figures from the Algorithmic Justice League of a 35% word error rate for African American voices against 19% for white voices. Different systems, different years, same direction.
A more recent preprint by Calbert Graham and Nathan Roll, "Evaluating OpenAI's Whisper ASR: Performance Analysis Across Diverse Accents and Speaker Traits" (Cambridge Open Engage, 7 November 2023, later published in JASA Express Letters), found Whisper performed better on North American English than on British and Irish English, that a speaker's first language and second-language proficiency were linked to the error rate, and that read speech was recognized better than conversation. Whisper is one of the most widely used open models, so this is a fair proxy for the engines behind many subtitle tools.
Named entities. A name has no context clue. When a speaker says "Thiruvananthapuram", "Oyelaran" or "Quezon City", the recognizer has to pick from words it has seen, and it picks the closest common word. The same happens with brand names, product names and handles. Accent makes it worse, because the sound the model is matching is already further from what it expects.
Code-switching. Most multilingual speakers drop words from another language into English sentences. "Send me the file by kal" or "that is the pangit version" is normal speech, but an English-only model has no entry for "kal" or "pangit" and writes an English word that sounds similar. The more switching in a clip, the more of these substitutions you will find.
YouTube's own help page on automatic captions, read on 18 October 2026, says the same thing in fewer words: captions "might misrepresent the spoken content", and it lists mispronunciations, accents, dialects and background noise as causes.
What the subtitle tools say about accents
Several generators mention accents on their pages, read on 18 October 2026. Maestra says its engine is "trained on regional pronunciation, accents, and dialects". HappyScribe's English subtitles page offers "AI subtitles in English and 120+ other languages and accents". Clipchamp says it works "no matter your dialect or accent". These are the vendors' own descriptions, and none of the pages publishes a per-accent test or an error rate by accent, so treat them as a statement of intent rather than a result. CapsAI does not publish an accuracy figure for any accent, and this post will not invent one.
The fix-the-draft checklist
Read the draft once with the audio playing. On a 60-second clip this takes three to four minutes. Fix in this order, because the first items are the ones viewers notice most.
- Names. People, places, companies, products, handles. Search the draft for every name you said and set the spelling the person or brand uses. Fix it everywhere, not just the first time.
- Numbers. Prices, dates, times, percentages, phone numbers, model numbers. Accented vowels push "fifteen" to "fifty" and "thirteen" to "thirty" more often than the reverse. Check every digit.
- Brand and product words. "Paytm", "Jollibee", "Safaricom", "M-Pesa" all have an English word a model will reach for instead. Decide on capitalization once and apply it.
- Words from your other language. Find each one and write it the way your audience writes it. If the clip has many, see the Hinglish section below.
- Repeated words and false starts. "We we need to" and "the the file" are usually the model being faithful to a stumble. Delete the repeat unless the stumble is the point.
- Homophones and sound-alikes. "their / there", "effect / affect", "won't / want", and the pairs your accent creates: for many Indian speakers "vest / west", for many Filipino speakers "fan / pan". Read these lines twice.
- Dropped word endings. "-ed" and "-s" endings are quiet in fast speech. "He ask me" and "two file" are typical and read as grammar errors on screen.
- Line breaks and punctuation. Break where you paused. Add the question mark where your voice rose. A correct word in the wrong line still reads badly.
In CapsAI every word, timing and style is editable in the browser before export. Click the line, retype, move on. The export is an MP4 with the captions burned in, so what you fixed is what viewers see.
Audio habits that cut errors before they happen
The draft gets better when the recording gets better, and these cost nothing.
- Record closer to the microphone. Phone at arm's length rather than across the room. Distance adds reverb, which is harder on accented speech than on anything else.
- Lower the background. Fans, traffic and a TV in the next room sit in the same frequency range as consonants. Close a door or move.
- Keep the pace even. You do not need to slow down, but avoid speeding up at the end of sentences, which is where endings get dropped.
- Say names twice in the video if they matter. Not for the model, for the viewer. The second mention is also your spelling check.
- Do not change your accent. Forced pronunciation produces hesitations and odd stress patterns that the model handles worse than your natural speech.
- Avoid music under speech during recording. Add it in the edit if you want it. The recognizer only hears the mix.
When to pick Hinglish instead of English
If you speak Indian English and the clip mixes in Hindi words, the question is how much. A sentence or two with "kal", "theek hai" or "matlab" is fine under the English setting; fix those few words by hand. When Hindi words appear in most sentences, pick Hinglish instead. The Hinglish caption generator treats the mix as one spoken language and writes it in Roman script with the English words kept as English, so "aaj we are launching the new plan" comes back as one line instead of an English guess at "aaj".
Hinglish is the only mixed-language setting. Speakers who mix English with Tagalog, Yoruba, Malay, Urdu or anything else should use the English setting and correct the borrowed words in the editor. If your video is mostly in Hindi, Tamil, Telugu, Bengali, Marathi, Gujarati, Kannada, Punjabi or Urdu, pick that language, for example through the Hindi subtitle generator; CapsAI captions what is spoken and does not translate. Indian language subtitle support compared lists which other tools handle these languages as spoken languages.
How to test your own accent in five minutes
- Choose a 60-second clip that is typical of how you speak on camera, with at least two names and a few numbers.
- Upload it to the English subtitle generator and pick English. A new account has 3 free minutes to start.
- Read the draft with the audio. Count the words you had to change, and note what kind they were.
- If most changes were names and numbers, that is normal for any tool and the checklist above covers it. If common words were wrong throughout, re-record with the audio tips and compare.
- Export the MP4 and watch it on a phone with the sound off. If you can follow the clip, the captions are doing their job.
This is also the honest way to compare tools. A vendor's accuracy figure was measured on their test audio, not on your voice. Your clip through the free minutes is the only number that describes your result.
What to do next
Caption your next English clip with the English subtitle generator, run the checklist on the draft, and export the MP4 with the captions burned in. If the clip mixes Hindi and English in most sentences, switch to the Hinglish caption generator instead. For a video in one of the other 10 Indian languages, the auto subtitle generator lists every spoken-language option. One-time packs start from $4 (₹49 in India, €4 in the euro area) after the 3 free minutes, and watermark-free export comes with every paid pack.
