BetterSpeak brings English language practice offline with Gemma 4 E2B
BetterSpeak brings English practice offline with Gemma 4 E2B
HubX uses Gemma 4 to deliver private, low-latency language tutoring on edge devices
HubX's English learning app, BetterSpeak, is a voice-based mobile tutor that helps users learn English in their native language. BetterSpeak’s goal is delivering the feeling of talking with a real tutor, regardless of your skill, location, or connectivity. And now, by running Gemma 4 on-device BetterSpeak offers uninterrupted conversation practice with no additional fees to the user.
Ensuring pedagogical standards for language learning
BetterSpeak supports a wide range of native languages, including Spanish, Chinese, French, German, Turkish, Portuguese, Russian, Japanese, and Arabic, and allows users to progress through beginner, intermediate, and advanced lessons, all using LLMs. Historically, BetterSpeak offered a high-quality conversational experience through online, cloud-based models. BetterSpeak aimed to expand that service and provide tutoring almost entirely offline.
The team introduced an on-device pipeline with Gemma 4 E2B as the central reasoning engine for their AI tutor, citing its ability to manage mixed-language inputs while ensuring consistent lesson states. Their tutor needed to provide natural grammar explanations, facilitate exercises, monitor progress, and offer contextual feedback—all while maintaining its persona. Long-form, deterministic prompts were used to define scenario-specific rules, ensuring the AI strictly follows pedagogical instructions across multiple turns.
Gemma 4 E2B proved capable of managing the complexity of providing native-language guidance alongside English learning materials, making expansion to other languages and global markets promising. Furthermore, the ability to run the model locally on the user’s device removes the financial burden of per-token API pricing.
Deploying conversational AI tutoring at the edge
Implementing a full conversational AI loop on edge devices involved overcoming memory and thermal limitations, particularly during the initial loading of the model. The BetterSpeak team enabled this offline capability, currently only available on iOS, by addressing the specific demands of mobile deployment:
- Payload optimization: The solution utilizes a ~2.9 GB package featuring the 4-bit quantized version of Gemma 4, balancing advanced reasoning with mobile storage limits.
- Hardware and stability: By targeting iPhone 15 Pro and newer models, the team managed thermal and memory constraints. Gemma maintained stable runtime performance during extended offline sessions during BetterSpeak testing.
- Strategic offloading: Offline mode is currently only available on iOS. Native Apple Dictation proved most effective for transcribing utterances that combined English with the user's native tongue, while Kokoro provided a lightweight text-to-speech engine.
Localized inference minimizes conversational lag, enabling a natural exchange where learners can ask for clarification or interrupt the tutor without losing momentum. Vocal inputs, text transcriptions, and the overall lesson logic are processed entirely on-device, enabling BetterSpeak to offer a privacy-preserving experience.
Robust offline learning unlocked
Deploying Gemma 4 E2B directly on edge devices allowed HubX to introduce a robust offline capability into BetterSpeak that complements their existing cloud-based services, allowing users to resume their experience from where they left off on their learning path.
As tooling and runtime support continue to mature, the team plans to keep evaluating native audio processing directly within the model, which would further minimize conversational lag and unlock more nuanced voice-to-voice interactions for language learners. Currently, the offline mode supports English, Turkish, and Spanish, and HubX is looking forward to expanding to more of their online-supported native languages, advancing their vision of helping more people speak better and learn faster with AI.
More from the Gemmaverse
CastFox adopts Gemma 3n to scale its interactive podcast experience
SEA-LION v4 is an open multimodal model trained for Southeast Asian languages and cultural contexts
Lettria fine-tunes Gemma 3 for reliable knowledge graph creation
