Hush Studio

Hush Studio is an Android app that turns written scripts into natural-sounding speech, entirely on the phone. You can narrate with one of six built-in voices or clone your own voice from a short recording. After a one-time model download, nothing is sent over the network: scripts, recordings and generated audio stay on the device.

It is built for long-form narration such as video voiceovers, audiobook drafts and podcasts, rather than instant one-line text-to-speech.

Features

• Offline speech generation using the VoxCPM 0.5B model, running on the phone's CPU.
• Voice cloning from 5 to 10 seconds of speech. You can record or import audio in any format the phone can decode. The app picks a clean section and writes the transcript offline. A cloned voice is saved as a reusable reference; no model is retrained.
• Six built-in voices made from openly licensed recordings (see Credits).
• Chunk studio. Each paragraph becomes its own chunk. You can generate, play, edit or regenerate any chunk without touching the others, and earlier takes are kept.
• Export a full narration as a 16 kHz mono WAV file.
• Library of projects with search, rename and delete. Unfinished projects are kept, so you can come back to them.
• Light and dark appearance that follows the system setting.

Requirements

Android	9 or later (API 28+)
CPU	64-bit ARM (arm64-v8a). Phones with ARMv8.2 dot-product and FP16 support (most from 2019 onward) use a faster engine; older 64-bit phones run a compatible engine automatically.
Memory	4 GB RAM minimum; 6 GB recommended. Generation peaks at about 1.7 GB.
Storage	About 650 MB free for the speech model, plus space for projects
Network	Needed once, to download the 500 MB speech model

Installation

1. Download the latest APK from the Releases page and install it. Android will ask you to allow installs from your browser or file manager.
2. Open the app and go to Settings > Speech Model to download the model, or import the voxcpm-0.5b-q4_k.gguf file if you already have it. The file is checked against a fixed SHA-256 hash before use.
3. Write or paste a script in Create, choose a voice and tap Generate.

Performance

Generation runs on the CPU and is slower than real time. On a Motorola Moto G71 (Snapdragon 695, 6 GB RAM), a three-paragraph script producing about 21 seconds of audio takes about 85 seconds with a built-in voice. That is roughly four times the length of the audio. Faster phones finish sooner. The Quality setting trades speed for clarity: Fast Preview uses 4 diffusion steps, Balanced 6 and Best Quality 10.

The model stays loaded for 90 seconds after a task, so regenerating a chunk soon afterwards skips the load. GPU acceleration is experimental and is not used by default.

Privacy

• No account, analytics or advertising.
• The only network request is the speech model download from Hugging Face.
• Scripts, voice samples, transcripts and narrations are stored in the app's private storage and are excluded from Android backups.
• Uninstalling the app deletes all voices and projects. Export anything you want to keep first.

Using voice cloning responsibly

Only clone your own voice, or a voice you have explicit permission to use. Do not use generated speech to impersonate or deceive people. When you publish generated audio, consider labelling it as AI-generated.

Building from source

Requirements: JDK 17 or later, Android SDK platform 36, NDK 28.2.13676358 and CMake 3.22.1.

./gradlew assembleDebug
./gradlew testDebugUnitTest lintDebug

The APK contains two builds of the speech engine: libvoxstudio_cpu_fast.so (ARMv8.2 with dot-product and FP16, built by the app module) and libvoxstudio_cpu.so (baseline ARMv8, built by the engine-compat module from the same sources). At startup the app checks the CPU and loads the fast engine when it is supported.

Release builds read signing details from an untracked keystore.properties file in the project root:

storeFile=/path/to/release.jks
storePassword=...
keyAlias=...
keyPassword=...

Without that file, assembleRelease produces an unsigned APK.

Device tests are run with adb shell am instrument after installing the app and test APKs with adb install -r. Avoid connectedAndroidTest on a phone with data you want to keep, because Gradle uninstalls the app afterwards. scripts/ contains helpers for provisioning test fixtures, running the generation benchmark, and fetching candidate voice clips.

Project layout

app/src/main/java/   Kotlin app: screens, studio, playback, import, generation jobs
app/src/main/cpp/    JNI bridge and the persistent model session for the native engine
engine-compat/       Builds the baseline ARMv8 engine from the same native sources
app/src/main/assets/ Built-in voices, offline speech-recognition model, licence notices
vendor/voxcpm/       VoxCPM.cpp inference runtime (ggml)
scripts/             Benchmark, test-fixture and voice-candidate tools
docs/                Technical notes, verification log and GPU experiments

Limitations

• English only.
• Output is 16 kHz mono. Write numbers and abbreviations out as words for the most reliable pronunciation.
• Generation is not real time on mid-range phones.
• Results vary between takes. Regenerate a chunk if a sentence comes out wrong.

Credits

• Speech model: OpenBMB VoxCPM (Apache-2.0), quantized GGUF from bluryar/VoxCPM-GGUF.
• Inference runtime: VoxCPM.cpp (Apache-2.0) and ggml (MIT).
• Offline transcription: Vosk with the small English model (Apache-2.0).
• Built-in voices:
  • Aria: LJ Speech, public domain.
  • Jenny (Dioco): Jenny TTS dataset, used under its licence, which requires this name.
  • Maya, Ethan and Cole: Hi-Fi TTS (NVIDIA), CC BY 4.0.
  • Priya: CSTR VCTK Corpus, University of Edinburgh, CC BY 4.0.

Full licence texts are included in the app under Settings > Open-Source Licenses and in app/src/main/assets/THIRD-PARTY-NOTICES.txt.

License

Hush Studio is licensed under the Apache License 2.0. Bundled third-party components, the speech model and the built-in voice recordings keep their own licences, listed above and in the notices file.