DhVaani-0.5 — MNN (Android / on-device)

DhVaani-0.5 ka INT8 ONNX version source repo Bbkblo/DhVaani-0.5-ONNX se MNN (.mnn) format mein convert kiya gaya. Target runtime: Android / on-device (MNN).

Files

File	Role	Size
`text_encoder_int8.mnn`	text → text_condition	6.3 MB
`fm_decoder_int8.mnn`	flow-matching decoder (t, x, cond) → v	126.8 MB
`vocoder_backbone.mnn`	mels → hidden (vocoder backbone)	52.1 MB
`mel_fb.npz`	mel filterbank	0.2 MB
`vocos_head.npz`	vocos head weights	2.1 MB
`tokens.txt`	vocabulary	~8 KB
`model.json`	config	~0.7 KB

Conversion

Converted with MNNConvert (built from MNN master), from ONNX opset 18 / IR v8:

MNNConvert -f ONNX --modelFile X.onnx --MNNModel out/X.mnn --bizCode X

Full log: convert.log. Helper script: convert_dhvani.sh (requires MNNConvert binary).

Inputs / Outputs (printed by converter)

text_encoder_int8.mnn — inputs: tokens, prompt_tokens, prompt_features_len, speed → output: text_condition
fm_decoder_int8.mnn — inputs: t, x, text_condition, speech_condition, guidance_scale → output: v
vocoder_backbone.mnn — input: mels → output: hidden

Note: converter printed "The model has subgraphs, please use MNN::Express::Module to run it" — so at runtime use MNN Express (MNN::Express::Module) for these graphs.

Android usage (short)

Add libMNN.so (arm64-v8a) to app/libs/jniLibs.
Load each .mnn file as an Express module and wire the pipeline: text encoder → flow-matching decoder → vocoder backbone (+ mel_fb.npz + vocos_head.npz).
Run the whole TTS graph to produce waveform.

📱 Complete Android Setup Guide

👉 ANDROID_SETUP.md (repo mein) — full guide:

Gradle/Maven mein MNN lib add karna
.mnn + assets app/src/main/assets mein daalna
MNNNetInstance Kotlin code (load models, run text_encoder, fm_decoder loop, vocoder)
Speed-up tips (numThread, FP16/Precision_Low, num_step audio, session reuse, warmup, etc.)

(Ye app to app integration ke liye ready hai.)

Source

Downloads last month: -; Downloads are not tracked for this model. How to track