Known for its voice memo features, Google’s Pixel Recorder app is mainly used for short memos, despite its potential for recording meetings and lectures.
However, today the Recorder app is becoming more widely used as the app’s developers have taken advantage of Gemini Nano, a powerful on-device Large Language Model (LLM). This integration introduces an AI-powered audio summarization feature that helps users find the right recordings more easily and quickly capture key points.
Earlier this month, Gemini Nano received an upgrade with the introduction of the new multimodality model. The Recorder app already takes advantage of this improvement to better summarize longer voice recordings and offers improved processing of grammar and nuance.
Recorder users now use the new AI-powered summary feature an average of 2 to 5 times per day, resulting in a 24% increase in the total number of recordings saved. This feature has significantly increased app engagement and user retention. The Recorder team also reported that feedback on the new feature has been positive, with many users appreciating the time savings.
The recorder’s developers initially explored a cloud-based solution and achieved high performance and quality, but to improve user accessibility and privacy, they decided to take an on-device approach. The development of Gemini Nano provided an ideal opportunity to create concise audio summaries while ensuring that data processing remained on the device.
On-device LLM is beneficial for users as it provides them with more privacy, less latency, and works anywhere as no internet is required.
– Kristi Bradford, Pixel’s core apps product manager, August 2024
To improve results, Recorder optimized the model using low-order rank adaptation (LoRA) with application-specific data. This enables Gemini Nano to consistently produce three-part summaries of transcripts that include speaker names, key insights, and themes.
Recorder developers have integrated the latest Gemini Nano model, known as Gemini Nano with Multimodality, to improve the summarization feature on Pixel 9 devices. Larger and more advanced than its predecessor on Pixel 8 devices, this new model offers improved performance, accuracy, and scalability. It also supports a larger number of tokens, allowing Recorder to summarize much longer transcripts than before. Currently, Gemini Nano with Multimodality is available exclusively on Pixels 9 Devices.
To take full advantage of the new model’s capabilities, the recorder developers expanded their dataset to include longer voice recordings, refined their evaluation methods, and created start-criteria metrics that focus on grammar and nuance.
Would you like to switch now to Pixels 9?
