AI speaking practice usually chains three parts: speech recognition that turns the learner's voice into text, a language model that writes a reply, and speech synthesis that reads the reply aloud. It gives learners unlimited low-pressure conversation turns, but its feedback is only as reliable as the recognition step, which makes more errors for some accents and can invent words. In Canada, practice with an app does not count toward immigration: Express Entry accepts only results from IRCC-approved tests such as CELPIP-General, IELTS General Training, and PTE Core.
Key takeaways
- A typical speaking tutor combines speech-to-text, a language model, and text-to-speech, and an error in the first stage carries into the others.
- Word error rate counts substitutions, deletions, and insertions divided by the number of words in the reference.
- A 2020 study of five commercial recognizers found an average word error rate of 0.35 for Black speakers and 0.19 for white speakers.
- Researchers found that about 1 percent of Whisper transcriptions contained entire phrases that were not in the audio.
- Express Entry requires results from approved tests, valid for less than two years, and CLB 7 in all four abilities for the Federal Skilled Worker Program.
What happens between a spoken sentence and the reply
When a learner speaks, the audio is first converted to text by an automatic speech recognition model. One widely studied example is OpenAI's Whisper, described in a paper posted on December 6, 2022, which was trained on 680,000 hours of multilingual audio paired with transcripts and reported accuracy approaching that of humans on standard benchmarks. The text then goes to a language model, which predicts a suitable reply. Finally a synthesis model voices the reply. Some apps add a correction layer, where the language model compares the learner's sentence with a more natural version and explains the difference.
The important consequence is that the tutor never hears the learner the way a teacher does. It works from a transcript. If the recognizer writes down a word the learner did not say, the correction layer will respond to the wrong sentence.
A worked example of word error rate
Recognition quality is measured as word error rate (WER): substitutions plus deletions plus insertions, divided by the number of words in what was actually said. The numbers below are an illustrative example, not data from a real product.
| Item | Text |
|---|---|
| What the learner said (7 words) | She would like a cup of coffee |
| What the recognizer wrote | She would like cup of copy |
| Errors | One deletion (a), one substitution (coffee became copy) |
| WER | 2 errors divided by 7 words, about 0.29 |
A WER of 0.29 means roughly three of every ten words are wrong, enough to make a tutor misjudge a sentence that was in fact fine. A low WER on one speaker group does not guarantee a low WER on another.
Where recognition goes wrong
A study in Proceedings of the National Academy of Sciences in 2020 tested speech recognition from Amazon, Apple, Google, IBM, and Microsoft on 19.8 hours of interviews with 42 white and 73 Black speakers. The average WER was 0.35 for Black speakers and 0.19 for white speakers. The best system, Microsoft's, still showed 0.27 against 0.15. The study concerned dialect and race rather than learners of English, but it shows that accuracy varies by how a person speaks, and learners with strong first-language accents may face a similar gap. No comparable published figure for language-learning apps was verified for this article.
A second risk is invention. In the paper Careless Whisper, presented in 2024, researchers found that about 1 percent of Whisper transcriptions contained entire hallucinated phrases, and about 38 percent of those hallucinations included explicit harms such as false authority or violence. Speakers with aphasia, whose speech has longer silences, were affected more. Long pauses, common in learners who are searching for a word, may create similar conditions, though the study did not test language learners.
What the tools do well and what they cannot judge
- Good fit: unlimited practice of everyday exchanges such as ordering, introductions, and phone-style small talk.
- Good fit: instant explanations of grammar in the learner's own sentence, which a learner can check against a reference grammar.
- Poor fit: judging exact pronunciation from a transcript alone, since a transcript contains no sounds.
- Poor fit: replacing feedback from a teacher on pragmatics, such as politeness or tone at work.
- Poor fit: predicting a test score, because approved tests are scored by their own methods.
Practice versus an approved test in Canada
For immigration, only specific tests count. The IRCC page on language testing for Express Entry lists CELPIP-General, IELTS General Training, and PTE Core as the approved English tests. Results must be less than two years old when a profile is completed and when a permanent residence application is submitted, and the test must be booked with an IRCC-approved agency. The Federal Skilled Worker Program asks for Canadian Language Benchmark (CLB) 7 in all four abilities for the first official language. Requirements differ by program and change, so the IRCC page is the reference. This article is general information, not immigration advice.
FAQ
Can AI correct pronunciation?
Only indirectly in tools that rely on a transcript: they can flag words that were misrecognized but cannot hear sounds the way a trained teacher can. Treat a flagged word as a prompt to check a dictionary recording.
Does AI speaking practice help for CELPIP or IELTS?
It can give extra turns of conversation, but the tests are scored separately and the IRCC page names the approved tests. No verified evidence on score gains from AI apps was found for this article.
Why does an app sometimes mishear a correct sentence?
Recognition accuracy varies with accent, background noise, and pauses. The 2020 PNAS study found large differences between speaker groups on five commercial systems.
Is it safe to speak personal details to an AI tutor?
Audio may be stored or processed by the provider, so the privacy policy of the specific app decides. Avoid sharing sensitive details and check how recordings are kept.
Related reading
- How AI supports English language learning
- Replika and AI conversation partners
- How AI is changing English communication
- Conversational AI chatbots from OpenAI
More in AI Basics.
Last verified: October 2026
Sources
- Radford A, et al., Whisper speech recognition paper (large-scale weak supervision), arXiv:2212.04356, December 6, 2022
- Koenecke A, et al., Racial disparities in automated speech recognition, Proceedings of the National Academy of Sciences, 2020
- Koenecke A, et al., Careless Whisper: Speech-to-Text Hallucination Harms, arXiv:2402.08021, 2024
- Language testing for Express Entry, Immigration, Refugees and Citizenship Canada, 2026