NimVoice

NimVoice is a fully offline local AI Text-to-Speech project built primarily in Nim and accelerated through the Intel OpenVINO C API.

Its core goal is simple:

• native executable
• no Python runtime for end users
• real Intel NPU acceleration
• practical local TTS on Windows AI PCs

Current Open-Source Status

The current open-source baseline focuses on the stable English pipeline.

• Runtime synthesis is fully offline.
• The application is written around Nim + OpenVINO C API integration.
• Intel NPU support is treated as a first-class requirement.
• The current public build defaults to --lang=en.
• Chinese synthesis is temporarily disabled in the open-source build while the multilingual path is being stabilized.

Why This Project Exists

• Why Nim? Nim compiles to native code and makes it practical to build a standalone app with explicit control over FFI, memory, and deployment.
• Why OpenVINO C API? It keeps the runtime boundary small and predictable while still unlocking CPU, GPU, and NPU backends.
• Why Intel NPU? NimVoice is designed for real AI PC deployment, not just CPU-only demos.
• Why no Python runtime? End users should not need pip, venv, PyTorch, or background scripting layers just to synthesize speech.

Runtime Architecture

To work around the NPU's static-shape requirements while preserving a usable TTS pipeline, NimVoice uses a heterogeneous CPU + NPU design:

             Nim (Native App)
              │
        Text Normalization (Pure Nim)
              │
              ▼
    ┌───────────────────────────┐
    │ SpeechT5 Frontend         │
    │ Backend: Intel CPU        │
    │ Autoregressive Decoder    │
    └─────────┬─────────────────┘
              │ Mel-Spectrogram
              ▼
    ┌───────────────────────────┐
    │ HiFi-GAN Vocoder          │
    │ Backend: Intel NPU        │
    │ Static Graph [1, 128, 80] │
    └─────────┬─────────────────┘
              │
       WAV Audio Output

Repository Layout

src/            Nim runtime and OpenVINO integration
models/         model files and compiled cache
electron-gui/   Electron + React frontend
benchmark/      benchmark outputs
build.cmd       main build entrypoint
run.cmd         CLI entrypoint
run-gui.cmd     GUI entrypoint

Requirements

• Windows 11 x64
• Nim 2.x
• MinGW / TDM-GCC toolchain
• Intel OpenVINO runtime DLLs
• Optional: Node.js for building the Electron GUI

Build

From the project root:

build.cmd

This script:

1. builds the Electron GUI if npm is available
2. compiles resources.rc
3. builds src/main.nim into bin/nimvoice.exe
4. copies OpenVINO runtime DLLs into bin/

Run

GUI

run-gui.cmd

CLI synthesis

run.cmd "Hello world from NimVoice."

The output WAV is written to bin/output.wav.

Benchmark

bin\nimvoice.exe bench --warm-runs=0

Benchmark results are written to benchmark/results.md and benchmark/results.json.

Performance Notes

NimVoice uses real OpenVINO PERF_COUNT profiling rather than placeholder counters. A dedicated C wrapper is used to bridge Windows x64 variadic ABI differences safely.

The project also supports compiled-model .blob caching for the NPU path, which significantly reduces repeated startup cost after the first compile.

Open-Source Boundaries

This repository contains the application layer:

• TTS pipeline logic
• GUI integration
• benchmark tooling
• model orchestration

The lower-level OpenVINO wrapper layer is being separated as the standalone Resonance library.

Attribution

The low-level Nim OpenVINO C API wrapper and binding work used by this project is attributed to Resonance:

https://github.com/AbyssGG/Resonance/tree/main

AI-Assisted Development

This project was developed with the assistance of Google Gemini 3.1 Pro.

AI was used as a development assistant for tasks such as code generation, debugging, documentation, API exploration, and development workflow support.

The project architecture, technical direction, integration decisions, testing, hardware validation, and final code review were performed by the project author.

AI assistance does not change the project's core technical implementation or its open-source licensing.

Related Files

• NOTICE.md
• TROUBLESHOOTING.md
• MODEL_LICENSES.md
• DEVLOG.md
• CONTRIBUTING.md
• README.zh-CN.md

License

NimVoice is licensed under the MIT License.

Third-party model and runtime licensing information remains documented separately in MODEL_LICENSES.md.