What the Research Says About Language Learning

Structured Pronunciation Practice Beats Casual Exposure by a Factor of 3

Here’s the number that should change how you learn a language: Cohen’s d = 0.89. That’s the overall effect size of explicit pronunciation instruction on second language learners’ accuracy, according to a meta-analysis of 86 studies. For context, that’s nearly three times the average effect size of most educational interventions (typically around 0.30–0.40). It means that if you spend 10 weeks doing structured pronunciation drills with feedback, you’ll improve your accent more than you would in 30 weeks of passive listening to podcasts. Most learners never do this — and that’s exactly why most learners plateau.

What the research actually shows

The meta-analysis on pronunciation instruction (Lee et al., 2015) pooled 86 experimental and quasi-experimental studies published between 1982 and 2013. The overall effect of d = 0.89 is considered large. But the details matter more than the average. When the researchers broke it down by type of instruction, the picture got sharper:

• Interventions that included corrective feedback (e.g., a teacher or app correcting your mispronunciation in real time) produced an effect size of d = 1.12 — even larger than the overall average.
• Interventions without feedback (e.g., just listening to native speakers or repeating without correction) produced d = 0.65 — still positive, but 42% smaller.
• Longer interventions (more than 4 weeks) outperformed shorter ones: d = 1.02 vs. d = 0.68 for programs under 4 weeks.
• Controlled outcome measures (reading word lists aloud) showed larger effects (d = 1.02) than spontaneous speech tasks (d = 0.72), meaning gains in controlled practice don’t automatically transfer to real conversation.

The meta-analysis included learners from a range of L1 backgrounds (mostly English, Mandarin, Japanese, and Spanish speakers learning English as an L2), with an average age of 22. Most studies ran for 6–12 weeks, with 2–3 sessions per week.

Now, contrast this with the broader language learning literature. A systematic review of replication studies in second language research (Porte & McManus, 2019) found that only 1 in every 400 articles is a replication, and the average replication takes over 6 years to appear. That means most of what we “know” about language learning techniques — from spaced repetition to immersion — rests on single studies that have never been independently verified. The pronunciation meta-analysis is one of the more robust findings in the field, precisely because it aggregates across many studies. But it’s an exception, not the rule.

The nuance most people miss

The pronunciation meta-analysis has a dirty secret: the effect on spontaneous speech (d = 0.72) is substantially smaller than on controlled tasks (d = 1.02). That gap of 0.30 standard deviations means that when you practice in a structured setting — repeating minimal pairs like “ship” vs. “sheep” — you get better at that specific task. But those gains don’t automatically show up when you’re trying to order coffee in real time. The transfer is real but incomplete.

This matters for self-experimenters because most people test themselves the same way they practice. If you drill pronunciation with an app that gives you a score on isolated words, and then you measure your progress with the same app, you’re measuring your ability to do the drill — not your ability to speak. The meta-analysis suggests you need to deliberately practice in conditions that mimic real conversation if you want real-world gains.

Also, the meta-analysis only included studies where the target language was English (for the most part). The phonemic inventory of your target language matters. If you’re learning a language with sounds that don’t exist in your native language (like tones in Mandarin or the uvular trill in French), the effect sizes might be different. The research doesn’t tell you that — it just didn’t test it.

Practical implications

• Practice pronunciation with corrective feedback, not just exposure. The meta-analysis found that feedback nearly doubled the effect size (d = 1.12 vs. d = 0.65). Use an app that gives you phoneme-level feedback (like ELSA Speak or Speechling), or record yourself and compare your production to a native speaker’s waveform. Do this for 15–20 minutes per session, 3–4 times per week.

• Extend your practice window to at least 8 weeks. Interventions under 4 weeks produced d = 0.68; those over 4 weeks produced d = 1.02. That’s a 50% improvement in effect size just from sticking with it. Plan for 8–12 weeks minimum, and don’t expect noticeable results in the first 2–3 weeks.

• Test yourself on spontaneous speech, not just drills. Because the transfer from controlled to spontaneous tasks is incomplete (d = 1.02 vs. d = 0.72), you need to measure what you actually care about. If your goal is better conversational accent, don’t measure progress with a word-list test. Record yourself describing a picture or telling a 1-minute story, and have a native speaker rate it blind.

Design your own experiment

What to test: Does structured pronunciation practice with corrective feedback improve your accent more than equal time spent on passive listening?

Intervention (A): 20 minutes per day, 4 days per week, for 8 weeks. Use a pronunciation app that provides phoneme-level feedback (e.g., ELSA Speak, or a tutor who corrects your minimal pairs). Focus on 3–5 sounds that are known to be difficult for speakers of your L1 learning your target language.

Control (B): 20 minutes per day, 4 days per week, for 8 weeks. Listen to native speaker audio (podcasts, TV shows, audiobooks) in your target language. No repetition, no feedback, no active production.

Design: Alternating weeks (A-B-A-B) or parallel weeks if you can recruit a friend to do the opposite condition. If you’re solo, do 4 weeks of A, then 4 weeks of B, with a 1-week washout in between.

What to measure: Before and after each 4-week block, record yourself doing two tasks: (1) reading aloud a list of 20 words containing your target sounds, and (2) describing a picture for 60 seconds (spontaneous speech). Have two native speakers rate each recording on a 1–7 scale for accentedness (1 = heavy accent, 7 = native-like). Average their ratings. Do not tell them which condition the recording came from.

Confound to watch for: The Hawthorne effect — you might improve just because you’re paying attention to your speech, regardless of the intervention. To control for this, keep a log of how much you speak the language outside the experiment (e.g., conversations with native speakers, language classes). If that number changes during the experiment, it’s a confound.

What a positive result looks like: Your accentedness rating improves by at least 1 point (on the 7-point scale) more in the feedback condition than in the passive listening condition. If the difference is smaller than 0.5 points, the intervention probably isn’t worth the extra effort over just listening more. If the difference is 1.5 points or more, you’ve found something that works — and you should keep doing it.