VoiceSift transcribes meetings, identifies speakers and distills keywords entirely on your phone โ the audio never leaves your device. AI-powered minutes available as an optional extra.
Transcription, speaker identification and keyword extraction all run on-device, end to end.
SenseVoice ASR runs entirely on your phone and supports Chinese, English, Japanese, Korean and Cantonese โ airplane mode included.
With the optional streaming model, subtitles appear while you record. Without it, transcription completes right after the recording ends.
Structured minutes from 14 scenario templates โ decisions, viewpoints and action items. Optional feature; see privacy notes.
Voiceprint recognition tells you who said what. Name a speaker once and every future meeting is auto-labeled.
Key terms are distilled from every meeting automatically, aggregated for management and searchable across all recordings.
Tap any transcript line to play the exact recording moment. Tap a keyword to jump straight to where it was said.
Save recordings as WAV, share via WhatsApp, WeChat, Telegram or any app โ or forward transcripts and minutes as plain text.
Custom vocabularies and deterministic word correction boost accuracy for legal, medical and technical terms.
All recordings, transcripts and profiles live in an encrypted on-device database. Recordings save to your Downloads folder on Android 10+ with zero permissions.
No analytics SDKs, no advertising identifiers, no telemetry, no account required.
Capture multi-speaker meetings with one tap.
Live subtitles with the optional streaming model โ or full transcription right after recording.
Keywords extracted and minutes generated the moment you stop.
Find any moment by keyword and replay the exact audio.
From lecture halls to boardrooms โ here is how people use VoiceSift every day.
"The professor talks so fast I can't keep up. VoiceSift turns the lecture into text โ I quickly locate the parts I didn't understand, read the transcript and replay the exact audio. Huge help."
"I record lectures and review afterwards โ tap any line to hear exactly how the professor explained it. Revision used to take hours; now it takes minutes."
"Meetings across time zones, action items everywhere. The AI minutes give me decisions and to-dos in a structured summary I can share with the team right away."
"Interviews transcribed word by word with speakers labeled โ I search a keyword and jump straight to the quote. The WAV export fits my editing workflow perfectly."
"Consultation records stay on my own device โ nothing uploaded. Keyword search finds the exact clause discussion in seconds. That's the privacy my clients expect."
"Academic talks are dense. I skim the extracted keywords first, then replay only the sections around them. I get more out of one talk than I used to from three."
VoiceSift was built privacy-first from day one โ not as a setting, but as an architecture. Here is exactly what happens to your data.
Speech recognition, speaker identification and keyword extraction all run locally. Your audio is processed on the chip, not in someone else's data center โ and is never uploaded.
Recordings and transcripts live in an encrypted local database โ no account, no sync, no cloud backup.
No analytics SDKs, no advertising identifiers, no telemetry.
Minutes generation is off by default. If you enable it, only the transcript text โ never audio โ is sent to the large-model provider you configure with your own API key. The app clearly discloses this and warns before sending sensitive recordings.
Free to use. Models download individually on demand, with mirror sources available.
Download APK| Model | Size | Purpose | Required |
|---|---|---|---|
| SenseVoice-Small ASR | 228 MB | Offline speech recognition | Core |
| Qwen2.5-0.5B LLM | 469 MB | Transcript polishing & keywords | Core |
| Zipformer-14M | 24 MB | Real-time subtitles (streaming) | Optional |
| 3D-Speaker | 27 MB | Speaker voiceprint | Optional |
| Qwen2.5-1.5B LLM | 1.07 GB | Higher-quality polishing (choose one of the two LLMs) | Optional |
Current version: v1.0.1
Everything you might want to know before recording your first meeting.
Chinese (Mandarin), English, Japanese, Korean and Cantonese. The interface itself is available in Simplified Chinese, Traditional Chinese, English, Japanese and Korean.
With the small optional streaming model (24 MB) installed, subtitles appear live while recording. Without it, the core SenseVoice model transcribes the full recording right after you stop โ same accuracy, just not live.
Minutes generation is optional and off by default. When enabled, only transcript text is sent to the large-model provider you configure with your own API key. Audio is never sent. The app clearly discloses this and reminds you to be careful with legally sensitive recordings.
Every model can be downloaded from multiple mirror sources, and downloads support resuming. Check your connection or switch networks, then retry from the in-app model manager.
Android 8.0+ on arm64-v8a devices. 2 GB of RAM is the minimum for the base configuration; 4 GB is recommended if you want the larger 1.5B model. Reserve roughly 1.5 GB of storage for models.