> ## Content Index > Fetch the complete content index at: https://community.mozilladatacollective.com/llms.txt > Use this file to discover other available public pages before exploring further. # From Scam Detection to Language Learning: How Builders Are Using Mozilla Data Collective Datasets - URL: https://community.mozilladatacollective.com/from-scam-detection-to-language-learning-how-builders-are-using-mozilla-data-collective-datasets/ - Published: 2026-11-11T00:21:35.000Z - Updated: 2026-11-11T00:21:35.000Z - Author: Mozilla Data Collective What happens after someone downloads a dataset from Mozilla Data Collective? For some AI builders, that data becomes part of a product already reaching users. For others, it plays an equally important role much earlier in development, helping them test an idea, improve a model or identify problems before a product ever launches. Access to diverse, high-quality datasets is important throughout that process. Beyond training AI models, real-world data helps developers understand whether their technology will actually work for the people it's intended to serve. Guardian and Sibúlo are two very different examples of how builders are using datasets available through Mozilla Data Collective to navigate those challenges. ## **Guardian: Building an AI Tool to Detect Phone Scams** Piotr Szmielew began developing Guardian after seeing the rise of phone scams in Poland, particularly those targeting older adults. Some scammers impersonate family members or bank representatives to convince victims to hand over money, and increasingly sophisticated tactics can include AI-generated voice clones. He wanted to explore whether an app could listen to phone conversations, identify suspicious patterns and warn users before they became victims. To build Guardian, Szmielew needed speech recognition technology that could understand Polish even when call quality was poor. He used Polish voice data available through Mozilla Data Collective, processed the recordings to simulate degraded cellular audio, and used them to fine-tune a speech recognition model. The result was a working prototype that could analyze calls using three approaches: identifying suspicious conversations through AI, flagging specific phrases commonly associated with scams, and attempting to detect AI-generated voices. Szmielew tested Guardian by calling his phone through a VoIP service and playing an actual recording of a scammer. The app successfully identified the suspicious conversation and displayed a warning on the screen. But getting the technology to work was only part of the challenge. Szmielew found that detecting AI-generated voices was particularly difficult, especially when running models directly on a mobile phone. And while the prototype worked, deploying it more broadly would require special Android permissions that typically aren't available to ordinary apps. Without support from a mobile carrier, Guardian couldn't move into production. For Szmielew, the project demonstrated both the potential of using speech datasets to develop protective technology and the practical challenges of turning an AI prototype into a consumer product. Access to Polish voice data helped him train and test a model for real-world calling conditions, while the development process revealed important limitations around mobile performance, deepfake detection and deployment. ## **Sibúlo: Making Mandarin Pronunciation Feedback More Reliable** Arnaud Liperini is developing Sibúlo, a Chinese language-learning app designed to help students not just recognize words, but understand and use them in conversation. Drawing on his own frustrations learning Chinese, Liperini wanted to create a tool that combines listening, reading, writing and pronunciation practice to help students retain what they've learned. Pronunciation was one of the biggest challenges. In Mandarin, the tone of a syllable can completely change its meaning, yet Liperini found that existing tools often provided limited feedback or overlooked important elements like tone. He decided to build his own analysis engine to help learners understand exactly where they were making mistakes and how to improve. To test whether the engine would work reliably outside a studio, Liperini turned to Common Voice's Chinese Scripted Speech dataset, available through Mozilla Data Collective. He selected nearly 1,000 recordings representing approximately 12,000 syllables, using community-validated clips from a broad range of speakers and recording conditions. Liperini expected recordings made on everyday devices to receive lower pronunciation scores than studio recordings. Instead, the overall scores were slightly better than anticipated, while a small number of syllables in otherwise validated recordings received unusually low scores. That meant the engine could incorrectly tell students they'd made a mistake even when their pronunciation was correct. "For a learner, being told 'wrong' after pronouncing a word well is the surest way to lose confidence in the tool, and in yourself," Liperini explained. The discovery led him to adjust how Sibúlo evaluates pronunciation, reducing the number of validated syllables incorrectly flagged to fewer than one in 100\. He also introduced a middle ground where the engine can withhold judgment rather than risk giving inaccurate feedback. Without access to Common Voice, Liperini would have struggled to collect such a diverse range of recordings on his own, and the issue might not have surfaced until students began using the app. "Without it, I would have discovered this problem with my first learners, by wrongly telling them they had made a mistake," he said. "Thanks to it, I fixed the problem before they could run into it." Sibúlo is now preparing for release, with a waiting list open at [sibulo.app](https://sibulo.app/?ref=community.mozilladatacollective.com). Liperini plans to expand the experience into stories, AI-powered conversations and opportunities for students to connect with teachers, with Common Voice continuing to support testing of future versions. **What Are You Building with Mozilla Data Collective?** We'd love to hear about it! Whether you're developing a new product, testing an idea or using our datasets to tackle a challenge, we're always interested in learning how builders are putting the data to work. Have a story to share? Get in touch with us.