VoiceSynth Neural Engine (Mobile / Edge ONNX)

High-performance, 100% offline neural voice cloning and text-to-speech engine optimized specifically for on-device mobile inference on modern Android flagships (Qualcomm Snapdragon 8 Elite / 8 Gen 3 and ARM64 NEON).

Published by latentdivergence.


Model Components

File Purpose Footprint Target Backend
tone_extract.onnx Extracts 256-d speaker timbre embedding from 3–5s audio 3.21 MB Qualcomm NNAPI / NPU (13 ms)
tone_color.onnx Normalizing-flow timbre adaptation network 149.91 MB Qualcomm NNAPI / NPU
tone_config.json Periodic Hann STFT audio DSP configuration (22,050 Hz) 838 B Engine Configuration
en_US-lessac-medium.onnx VITS neural acoustic base speech synthesizer 60.27 MB Multi-threaded CPU XNNPACK
en_US-lessac-medium.onnx.json Phoneme dictionary, IPA mapping, length scales 4.88 KB Engine Configuration
model_manifest.json Production delivery manifest with SHA-256 signatures 1.8 KB In-App Downloader & Verifier

On-Device Physical Benchmarks

Measured on physical hardware running Android 16 (API 36):

1. OnePlus 13 (Snapdragon 8 Elite / SM8750 "sun")

  • Speaker Embedding Extraction: 13 ms
  • Model Cold Start: 1,172 ms
  • Model Warm Start: 633 ms
  • Real-Time Factor (RTF): 0.550x (1.8x faster than real-time)
  • Active Synthesis RAM: 91.7 MB
  • Thermal Status: Nominal (Cool)

2. OnePlus 12 (Snapdragon 8 Gen 3 / SM8650 "pineapple")

  • Speaker Embedding Extraction: 25 ms
  • Model Cold Start: 1,549 ms
  • Model Warm Start: 588 ms
  • Real-Time Factor (RTF): 0.499x (2.01x faster than real-time)
  • Active Synthesis RAM: 71.1 MB
  • Thermal Status: Nominal (Cool)

In-App Model Distribution

Production Android applications can keep their base APK lightweight (~15 MB) by fetching this model bundle on first launch using the included model_manifest.json:

https://huggingface.co/latentdivergence/voicesynth-neural-engine/resolve/main/model_manifest.json

All models are served via high-speed global CDN with SHA-256 cryptographic verification.


License

This model suite is licensed under the MIT License. Commercially unencumbered for embedded, on-device, and edge applications.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support