RVC Android

English | 한국어

Voice conversion that stays on your phone — An Android app that swaps your voice using voice-changer-compatible RVC models.

Unofficial community tool. Not affiliated with the voice-changer project.

   

When you'd want this

Sometimes you just want to change a voice with the phone in your hand — no PC, no network.

• Turn a recording into someone else's voice — record a short clip from the mic or pick an audio file, and convert it with an RVC model you already have.
• Keep your audio off the wire — conversion and encoding both happen on the device. Nothing is uploaded, so it works in airplane mode.
• Swap models whenever you like — there's no voice baked into the app. Point it at whatever RVC ONNX model you want, right then and there.

What it does

• Changes your voice — feed it a mic recording or an audio file (up to 60s) and it converts the speaker with your RVC model.
• Exports the format you want — save as WAV, MP3, AAC, M4A, FLAC, or OGG.
• Lets you pick the models — choose the Synth, HuBERT, and RMVPE models with a file picker. Recording straight from the mic works too.
• Shows you the input first — files over 60 seconds are rejected up front, and the input waveform is drawn as a thumbnail.
• Plays the result right away — when conversion finishes, a modal pops up so you can listen on the spot, then export with "Save as…".
• Keeps your recent results — past conversions stay in a history card, so you can reopen a result you dismissed and play or save it again.
• Stays fast with the same models — a loaded model is kept warm in memory, so re-converting with the same combination skips reopening the model.

What you need

• Device: arm64-v8a, Android 12+ (minSdk 31, targetSdk 36)
• Permission: RECORD_AUDIO, only when you record from the mic (not needed for file input)
• Three models — all exported as ONNX from the voice-changer tooling. All three slots must be filled before a conversion starts.

Slot	Model	Good to know
Synth	RVC synthesizer ONNX	Must be the export2onnx.py output format from voice-changer. The model has to carry a custom_metadata_props["metadata"] JSON (samplingRate / f0 / embChannels / embedder / embOutputLayer / useFinalProj). Without it, the model is rejected with "synth has no embedded metadata".
HuBERT	ContentVec / HuBERT embedder	Like voice-changer's content_vec_500.onnx, exposing the unit12 (768d, v2), units9 (256d, v1), and unit12s outputs. Which one is used is chosen automatically from the Synth metadata.
RMVPE	Pitch extractor	Takes waveform[1,N] f32 and threshold[1] f32 inputs. Required for f0 models.

You get the Synth model by hitting export to onnx in the voice-changer desktop client — that's what bakes the required metadata into the file.

Getting started

To build and install it yourself:

./gradlew :app:assembleDebug

Once it's installed:

1. Open the app and pick the Synth · HuBERT · RMVPE models with the file picker.
2. Import an audio file or record from the mic (up to 60s).
3. If you like, adjust the pitch (f0UpKey) and Speaker ID.
4. Tap Convert, then listen to the result and export it with Save as….

Good to know

• It all runs on the device, offline — no account, no cloud, and your audio never leaves the phone.
• Input is capped at 60 seconds; longer files are filtered out before conversion.
• Models can be large, so the app runs with largeHeap and streams the model file to cache to read it via mmap (avoiding Java-heap OOM).

Under the hood

The conversion pipeline (see inference/RvcPipeline.kt):

1. Decode input → resample to 16 kHz mono (linear interpolation)
2. HuBERT/ContentVec → feats[1, T, C] (50 fps)
3. RMVPE → pitchf + mel f0_coarse quantization that matches voice-changer bit-for-bit, with the f0UpKey semitone shift applied
4. Embeddings upsampled 50 fps → 100 fps, 2× nearest (compatible with PyTorch F.interpolate(scale_factor=2))
5. Synthesizer → audio, clipped to [-1, 1]
6. Encode with ffmpeg-kit-audio (or in-process WAV) → SAF "Save as…"

Module layout:

app/src/main/java/com/ouor/rvcandroid/
├── MainActivity.kt
├── audio/         # decode/encode, resample, recording, preview player, history LRU
├── inference/     # ORT session cache, HuBERT/RMVPE/Synth wrappers, metadata parser
└── ui/            # Compose screen + ConversionViewModel (StateFlow-based)

Tech stack:

• Kotlin + Jetpack Compose / Material3 — UI
• ONNX Runtime Android — inference
• ffmpeg-kit-audio (community fork, LGPL) — MP3 / AAC / M4A / FLAC / OGG codecs
• AndroidX Media3 ExoPlayer — previewing the result