Build apps that record, transcribe, summarize, rewrite, and replay voice memos hands-free. Wake word activation, voice commands, real-time transcription, LLM summarization, and spoken playback. All on-device. No audio or text data ever leaves the device.
On-device voice memo assistant combines Porcupine Wake Word, Rhino Speech-to-Intent, Cheetah Streaming Speech-to-Text, picoLLM, and Orca Streaming Text-to-Speech into a single local pipeline that records, summarizes, rewrites, and replays voice notes. Porcupine listens for the wake word, Rhino captures intents such as start, summarize, rewrite, or read the memo. Cheetah transcribes the dictation. picoLLM produces the summary or rewrite locally. Orca speaks playback. Memo content never leaves the device.
Porcupine Wake Word is lightweight, accurate, customizable, and production-ready. It enables an always-listening wake word with very low CPU and battery cost, which is critical for a voice memo app that lives in the background and needs to be activated without draining the battery of mobile devices. With Porcupine, enterprises can train branded wake words and always-listening commands in seconds and deploy them across platforms. Industry-leading accuracy that fits on MCUs.
Rhino Speech-to-Intent maps spoken commands, such as "Start memo" and "Read memo," directly to actions with no intermediate transcript. Most voice command systems run a two-step pipeline: speech-to-text converts audio to a transcript, then a separate NLU model parses that transcript for intent. Every step accumulates error and compounds latency. Rhino achieves 6× higher accuracy than STT-plus-cloud-NLU stacks offered by Big Tech, even in noisy environments.
Cheetah transcribes the memo dictation in real time as the user speaks. Cheetah matches or beats cloud STT API accuracy and supports custom vocabulary for industry jargon and proper nouns, which is important for memos that include client names, drug names, case numbers, or part numbers. Cheetah emits words at 590 ms average latency, typically one word behind the speaker, and requires less compute than any other local engine tested. The result: no tradeoff on accuracy, latency, or privacy.
picoCompression quantizes language models to run on phones, browsers, and embedded boards with no cloud dependency while preserving task accuracy, making it well-suited for memo summarization and rewriting. picoLLM's minimal memory footprint allows the model to run alongside Cheetah, Rhino, and Orca in a single session without introducing network latency or privacy exposure.
Orca reads back the memo, the summary, or the rewritten version with a natural-sounding voice. Orca Streaming TTS is built for real-time LLM applications. It starts reading LLM responses as soon as LLMs create a meaningful word or phrase, achieving first-token-to-speech latency of 130 ms. Orca's 29 MB peak memory allows it to run alongside Cheetah, picoLLM, Porcupine, and Rhino with no performance issues. Most high-quality TTS engines require hundreds of megabytes of RAM, which limits where they can deploy.
Sales reps walk out of a meeting and record their notes into their phones or tablets on the way to another meeting or home. picoLLM structures the transcription into a summary with next steps, attendees, and follow-ups. The captured note can be pushed into the CRM. Customer voice and prospect details never leave the device.
Doctors, nurses, and therapists can dictate patient encounter notes between visits, reducing time spent on documentation. picoLLM summarizes them into a SOAP-friendly structure. No audio or transcript is sent to a cloud, which removes BAA requirements under HIPAA. Porcupine, Cheetah, and Rhino all support custom medical vocabulary.
Lawyers, journalists, and financial advisors capture interviews containing sensitive data that shouldn't be shared with third parties. Yet, besides Big Tech, newer companies like Otter.ai have hit the news and courtrooms for questionable privacy practices. On-device AI is the only way to keep voice memos private.
Service technicians inspect assets and capture issues on manufacturing lines and in the field, where connectivity is unreliable, and content has commercial-sensitivity or source-protection requirements. On-device processing ensures notes are captured securely and accurately without depending on a network connection.
A complete working recipe in Python. Open-source on GitHub. Runs 100% on-device.