VoiceAI Android model bundle (offline English → Hindi voice translation)
Everything the VoiceAI Android app downloads on first launch so it can work fully offline:
English speech → English text → Hindi text → Hindi speech. The app fetches
https://huggingface.co/Kg1729/voiceai-android-models/resolve/main/<path> for every file in
models.json (size + SHA-256 per file) and verifies each checksum.
| Path | Model | Runtime on the phone | License |
|---|---|---|---|
vad/silero_vad.onnx |
Silero VAD | sherpa-onnx | MIT |
stt/moonshine-base/ |
Moonshine base (English), int8, sherpa-onnx export | sherpa-onnx | MIT |
mt/indictrans2/ |
IndicTrans2 en→indic distilled 200M by AI4Bharat, our own int8 ONNX export (encoder + one-step decoder with a KV cache, de-duplicated vocabulary) + a SentencePiece tokenizer model for onnxruntime-extensions | ONNX Runtime | MIT |
tts/piper-hi/ (espeak-ng data as espeak-ng-data.zip, unpacked by the app) |
Piper voice hi_IN-rohan-medium (int8, sherpa-onnx export) + espeak-ng data |
sherpa-onnx | Voice trained on IIT Madras IndicTTS "Hindi Mono Male" data: see its license. espeak-ng data: GPL-3.0 |
These are redistributed unchanged from their public releases (Silero VAD, Moonshine and Piper via
sherpa-onnx releases, Piper also on
rhasspy/piper-voices), except IndicTrans2, which is converted
by edge/export_indictrans2.py + edge/dedupe_indictrans2.py in the VoiceAI project.
Note on the Hindi voice: the IndicTTS data license could not be verified when this bundle was published. If you reuse this bundle, check it for your use case, especially commercial use.
Quality (measured in the VoiceAI project)
- Speech recognition: WER 7.2 % on LibriSpeech (dummy set), vs 5.6 % for Whisper small.en
- Translation: chrF++ 50.4 on IN22-Conv (FLORES 59.9), the same on a Samsung Galaxy F15 (Android) and on the laptop
- Hindi speech: intelligibility CER 15.8 % on the phone (Whisper-small judge)
Total: 16 files, ~508 MB (the 355 espeak-ng files ship as one zip: fewer requests, no rate limits).