Last modified 25.07.2025
Drops utilizes artificial intelligence (AI) in its mission to make language learning engaging and effective, and this document describes how we safely use AI to support our language learners.
To help learners improve their pronunciation skills, Drops uses AI models and voice intelligence to assess the end user's ability to pronounce words in the language they are learning. The system compares their pronunciation with recordings of the correct pronunciation, providing immediate feedback. This feature is available to all Drops users.
Drops uses open-source AI models and does not build its own large language models (LLMs). No data from the Drops platform is used to train or refine any models.
Drops' main partner and LLM provider for its AI features is Hugging Face, using the Llama model (open source by Meta). Drops has integrated with Hugging Face through our backend to avoid unknown information from the browser or client being sent to Hugging Face.
Data inserted into AI features in Drops is sent to Hugging Face for processing and to generate AI output. The information Drops sends to the APIs of Hugging Face is what we call the prompt. Drops does not knowingly provide any user information, personal data or previous prompts to Hugging Face in providing its AI features and advises users against providing the same in prompts.
While AI speech recognition can enhance your language learning experience, they sometimes provide inaccurate responses or assessments. Please understand that pronunciation feedbac are generated by AI and may not always be perfect.
Please note that you as the user of Drops utilizing AI-powered features remain responsible for the content you provide during your interaction with the Drops platform, as well as its compliance with Drops' terms and policies.
You should not:
Key technical information related to the features:
When a user engages with Speak Practice, their voice is recorded locally and securely sent to Drops’ backend for processing. The audio is transcribed into text using various open-source automatic speech recognition (ASR) models hosted via Hugging Face. The resulting transcription is then compared against the target phrase using internal logic to generate a pronunciation similarity score. No personal data is included, the audio is not stored after processing, and no user recordings are used to train or improve the models.
The user may choose to not engage with the Speak Practice feature by skipping it when presented during gameplay. This will delay it from showing it for one hour.
Should you have any questions regarding these features, please contact our support team (sup@languagedrops.com).
