Talk to her
Hold to talk. She transcribes while you speak, then answers out loud with lip sync and subtitles.
PocketCompanion is a talking, lip-synced AI avatar for Android. Chat by voice, text or photo with a local LLM that runs entirely on your phone. No cloud, no account, no subscription.
Real screen recordings on a mid-range phone. Turn your sound on. The wait while she looks at a photo is fast-forwarded (it says so on screen) and an avatar-loading pause is cut; everything else is real time.
Hold to talk. She transcribes while you speak, then answers out loud with lip sync and subtitles.
Attach a picture and she describes what she sees, still fully offline.
Pick one of 11 avatars and 14 voices, and write her personality yourself. Here: a pirate captain.
A full voice assistant stack, an avatar engine and a language model, squeezed into one private app.
11 realistic digital humans by duix.com whose mouths move with every word.
Gemma 4 E2B on llama.cpp. Her reply starts streaming in about half a second, no server involved.
NVIDIA Nemotron streaming speech recognition shows what she hears as you talk.
14 Kokoro voices with American, British, Japanese and Mandarin accents.
Share a picture and she comments on it, thanks to Gemma 4 vision.
Word-by-word captions synced to her voice, like your favorite short videos.
Rewrite her instructions in Settings to change who she is and how she talks.
No account, no analytics, no backend. Downloads models once, then works offline.
Four open models and one avatar engine, chained together on the phone's CPU.
Hold the mic, type, or attach a picture.
Streaming recognition in 560 ms chunks via sherpa-onnx.
A local llama-server only the app can reach.
Each sentence becomes natural speech.
Mouth shapes computed from the audio, plus subtitles.
Measured on a POCO X8 Pro (MediaTek MT6899, 12 GB RAM), CPU only.
Free and open source. Install the APK, tap Download and start, and the app fetches about 5 GB of models once (use Wi-Fi). After that, no internet needed.
Yes. The first start downloads the models (about 5 GB). After that, the language model, speech recognition, voice and avatar all run on your phone, and you can use it in airplane mode.
It covers similar ground, a companion with a personality you can talk to, but nothing is sent to a server and there is no account. It does not yet remember past conversations.
64-bit Android 10+ phones with at least 8 GB of RAM; 12 GB is recommended. It was developed on a POCO X8 Pro. Recent mid-range and flagship phones should work; older ones will be slower.
No, it runs on the CPU. GPU and NPU acceleration are on the roadmap.
The default model is an abliterated build of Gemma 4, meaning its refusals were removed, so it answers more freely than a filtered assistant. Her instructions in Settings shape her personality and what she talks about.
Her instructions are editable in Settings. Any llama.cpp-compatible model can be used by changing a few constants and rebuilding; see the README.
Yes. The app code is MIT licensed. The avatar engine is duix.com's SDK under the duix.com Community License.