android_onnx_runtime

Android ONNX Runtime Example

This example is a small Android project that runs the MOSS-TTS-Nano ONNX Runtime path on device and writes a WAV file.

It intentionally stays minimal:

• no product UI
• no server calls
• no model files committed to git
• no app-specific business logic

The demo synthesizes two pre-tokenized prompts so the ONNX path can be tested without adding a large SentencePiece JNI dependency to the first Android example.

Model Files

Download the ONNX assets:

huggingface-cli download OpenMOSS-Team/MOSS-TTS-Nano-100M-ONNX \
  --local-dir MOSS-TTS-Nano-100M-ONNX

huggingface-cli download OpenMOSS-Team/MOSS-Audio-Tokenizer-Nano-ONNX \
  --local-dir MOSS-Audio-Tokenizer-Nano-ONNX

Copy both directories to the app external files directory:

Android/data/com.openmoss.ttsnano.onnxruntime/files/moss_tts_onnx/
  MOSS-TTS-Nano-100M-ONNX/
    browser_poc_manifest.json
    tts_browser_onnx_meta.json
    tokenizer.model
    moss_tts_prefill.onnx
    moss_tts_decode_step.onnx
    moss_tts_local_fixed_sampled_frame.onnx
    *.data
  MOSS-Audio-Tokenizer-Nano-ONNX/
    codec_browser_onnx_meta.json
    moss_audio_tokenizer_decode_full.onnx
    *.data

You can use Android Studio Device Explorer or adb push after launching the app once:

adb shell mkdir -p \
  /sdcard/Android/data/com.openmoss.ttsnano.onnxruntime/files/moss_tts_onnx/

adb push MOSS-TTS-Nano-100M-ONNX \
  /sdcard/Android/data/com.openmoss.ttsnano.onnxruntime/files/moss_tts_onnx/

adb push MOSS-Audio-Tokenizer-Nano-ONNX \
  /sdcard/Android/data/com.openmoss.ttsnano.onnxruntime/files/moss_tts_onnx/

Run

Open examples/android_onnx_runtime in Android Studio, connect a device, and run the app configuration.

Tap either demo button. The app writes a WAV file to its cache directory and prints the output path on screen.

The sample uses:

• moss_tts_prefill.onnx
• moss_tts_decode_step.onnx
• moss_tts_local_fixed_sampled_frame.onnx
• moss_audio_tokenizer_decode_full.onnx

Custom Text

For custom text input, tokenize with tokenizer.model using the same SentencePiece model used by the Python ONNX runtime, then pass the resulting token ids into MossOnnxDemoEngine.synthesize.

For a production Android app, add one of the following tokenizer paths:

• a small SentencePiece JNI wrapper
• a pre-tokenization service or build step
• another Android-compatible SentencePiece implementation

The ONNX Runtime code is independent from the tokenizer as long as it receives the correct IntArray token ids.

Notes

• Start with cpuThreads = 2 or cpuThreads = 4; device thermal behavior varies.
• The demo caps generation to maxFrames = 160 for faster smoke testing.
• The decoded ONNX codec output is stereo; this example averages channels and writes a mono WAV for simplicity.
• Keep the model files outside the APK for local testing. Bundling them into app assets is possible but increases APK size substantially.