vosk-tts-ru-0.9-prepared

Vosk-TTS ru 0.9-multi — synthesizer prepared for fast loading on Android

This is one file from vosk-model-tts-ru-0.9-multi, passed through onnxruntime.quantization.shape_inference.quant_pre_process (symbolic shape inference + graph simplification).

The weights are not touched. No quantization, no precision loss. Only the graph structure is cleaned up.

Why

The original model.onnx takes 25 seconds to load on an Android phone (Snapdragon, 8 cores) — and that is not graph optimization: it is equally slow with optimization disabled. The prepared file loads in 1 second.

Measured on device, ONNX Runtime 1.24.3, best of three runs:

Quantization turned out unnecessary: the prepared full-precision file is just as fast to load and loses nothing.

What is here

The language model needed no preparation — it already loads in 0.9 s. It is mirrored here only so that everything sits in one place; it is byte-identical to the file inside the official archive.

Not here: the dictionary file (101 MB). It is a plain text pronunciation dictionary and is better converted to a binary format for mobile use.

Everything original is at https://alphacephei.com/vosk/models.

Как это сделано

from onnxruntime.quantization.shape_inference import quant_pre_process
quant_pre_process("model.onnx", "model_prep.onnx", skip_symbolic_shape=False)

Файл получен из официальной модели Vosk-TTS без изменения весов — только приведение графа в порядок. Исходная модель грузится на телефоне 25 секунд на любом уровне оптимизации, включая отключённую; подготовленная — за одну.

Credits

Original model and all training work: Alpha Cephei, alphacep/vosk-tts. Licensed under Apache 2.0, same license applies here.