Problem Statement  

How might we improve Duolingo’s AI to better recognize and adapt to African accents for a more inclusive language learning experience?

Problem Background  

Since the beginning of the pandemic in 2020, more users from diverse backgrounds, including Africa, turned to Duolingo for language learning. However, many users with African accents reported that the AI frequently misinterpreted their speech, leading to frustration and incorrect feedback. This challenge highlights a gap in Duolingo’s AI training, which is primarily focused on American, British, and European accents.

Research Insights

User Pain Points

• Users with African accents frequently experience misinterpretations by Duolingo’s AI, impacting learning progress and satisfaction.
• Misinterpretation often leads to incorrect feedback, causing frustration and reducing user confidence in their speaking abilities.
• Lack of transparency in feedback makes it difficult for users to understand if errors are due to pronunciation or AI recognition limitations.

Supporting Data

^{85% of surveyed users indicated that AI’s inability to recognize their accents accurately was their biggest pain point. Analysis of user interactions showed that 70% of mistakes were linked to accent misinterpretation, not actual pronunciation errors.}

Landing on the Solution

Based on our target users’ pain points, we focused on enhancing AI’s ability to recognize diverse African accents accurately. This includes expanding training data to incorporate African-accented speech and improving transparency in AI feedback.

Explanation of Solution

After presenting our prototype to users again, we discovered that their main concern was not about pronunciation corrections but rather how accurately the AI understood their speech. Users appreciated the new AI confidence scores, which clarified whether errors were due to misinterpretation or actual pronunciation issues. This feedback confirmed that our focus on improving recognition accuracy and providing transparent feedback was on the right track. By addressing these concerns, we were able to enhance user trust and satisfaction with the AI’s performance

Mockups

Figma 

Future Steps

This is what we learned from speaking to customers: Users want AI to focus more on accurately recognizing diverse accents rather than offering pronunciation corrections. Many users also expressed a need for clearer feedback on whether errors were due to recognition limitations or actual mistakes in pronunciation.

Possible additional problems to address:

• Limited support for tonal languages: Users learning languages like Mandarin or Yoruba indicated that the AI struggles with tone recognition.
• Inconsistent feedback across accents: Some users reported that the AI's accuracy varies significantly between different African accents, highlighting the need for more balanced training data.
• Lack of real-world conversation practice: Users want AI-driven scenarios that simulate practical conversations beyond scripted dialogues.‍‍