Repeat one model recording twenty times and you may simply memorise that speaker saying that word. Change the voice or the vocabulary, and *ship* and *sheep* become indistinguishable again.
High Variability Phonetic Training (HVPT) addresses this problem directly. The target contrast appears in recordings from multiple talkers, in different words and phonetic contexts. You make a decision on every trial and receive immediate feedback.
Why variety matters for perception
A 2025 meta-analysis brought together 79 HVPT studies. It found medium-to-large overall effects on second-language speech perception, evidence of retention and some generalisation to novel stimuli. Effects varied with task and test design, total training time, target sounds and the number of talkers, among other factors.
The key phrase is speech perception. The evidence primarily supports learning to discriminate and identify sounds; it does not establish that an entire unscripted conversation will become fluent. Some studies also report production gains, but production still needs deliberate practice.
HVPT is most relevant when you know a word in print but mishear it in speech, when two sounds feel identical, or when you understand one familiar teacher but not a new voice.
Build a ship/sheep listening set
Choose one meaningful contrast
Start with a pair you actually confuse, such as *ship/sheep*, *full/fool* or *bad/bed*. Do not put ten minimal-pair families into one beginner session.
Prepare six to ten familiar words and confirm their meanings. The task is sound discrimination, not learning a new vocabulary list.
Collect several trustworthy models
Use reliable dictionary or learning audio from at least three speakers. Include different words and, later, short sentences. Repeating one synthetic voice does not create the variability you are aiming for.
Play one item without showing the text and choose *ship* or *sheep*. Then reveal the answer. If your tools cannot randomise automatically, number the clips and have a partner select them, or use a playlist whose filenames do not reveal the answers.
Get feedback after every choice
Do not listen ten times before committing to an answer. Choose, check and replay the relevant contrast if necessary.
Record the direction of the error: hearing *ship* as *sheep* may not be equally likely as hearing *sheep* as *ship*. The pattern helps decide what to review.
Move into production as a separate step
Put the words into short sentences: *The ship is near the shore.* and *I saw a sheep on the hill.* Listen to a model, then record your own version.
Do not exaggerate the mouth movement or stretch every vowel. Preserve the distinction while keeping the sentence rhythm understandable.
An eight-minute session
- Two minutes: review yesterday's most difficult contrast without text.
- Three minutes: try a new set or a new speaker, with feedback on each answer.
- Two minutes: put four frequent errors into sentences and speak after models.
- One minute: mix in old items to check that you have not memorised today's order.
On the next day, test a few unfamiliar voices before repeating the same list. Recognition with a new speaker and new words is more informative than a perfect score on a memorised playlist.
Three common misunderstandings
Listening to varied audio is not automatically HVPT. A podcast has many voices, but without a focused contrast, an explicit choice and immediate feedback, it is general listening input rather than the complete training procedure.
Perception gains do not solve every production issue. Hearing a contrast is a foundation. Articulation, duration and maintaining the difference inside a sentence require their own work.
An app judgement is not infallible. Recognition is affected by noise, accent and context. Use it as one source of feedback and return to the original audio for important comparisons.
How MirrorSay complements the listening work
Once you can hear the contrast, say short sentences containing those sounds to MirrorSay, compare the natural model and practise the whole rhythm. MirrorSay's word-level timing comparison is not a phoneme judge; it cannot certify every vowel. It is useful for putting sounds you can identify into connected, understandable speech.
For tongue placement, see How to Pronounce TH. If individual words are clear but the whole sentence is hard to parse, continue with Connected Speech and Linking.
References
- Sakai, M., & Moorman, C. (2025). High variability phonetic training (HVPT): A meta-analysis of L2 perceptual training studies. *Studies in Second Language Acquisition*, 47(3), 794–827. Open-access article
- Sakai, M., & Moorman, C. (2018). Can perception training improve the production of second language phonemes? A meta-analytic review of 25 years of perception training research. *Applied Psycholinguistics*, 39(1), 187–224. DOI
*MirrorSay is a free AI speaking-practice app: say a sentence, see a natural way to express it, then shadow the model. No sign-up, and speech is processed on-device whenever possible.*