ParSiGo (persian_tts) is a Persian text-to-speech project by nimaone that runs fully offline on CPU, with no GPU and no internet connection at runtime, and supports voice cloning from a short reference clip. It takes Persian text directly, handles phonemisation and synthesis itself, and writes WAV output at a 24 kHz sample rate. The synthesis model is mehdi-hf/pocket-tts-farsi-v2, a Persian model published by mallahyari and built on Kyutai's MIT-licensed pocket-tts library; this repository packages it and adds an ONNX export path, a Persian-to-phoneme G2P stage (mehdi-hf's repackaging of Homo-GE2PE-Persian by Elnaz Rahmati and colleagues, MIT), and a FastAPI web demo. Two execution paths are provided: a pure ONNX Runtime engine without torch, listed at roughly 200MB of installed dependencies on Windows (onnxruntime, numpy, scipy, soundfile, sentencepiece; scipy is only used to resample reference audio but is imported unconditionally for now) versus roughly 1.2GB for the torch reference path on the same system. On a 2.8s clip the README reports 3283ms for ONNX against 3423ms for torch, with a waveform difference of 7.5e-4 from the torch output. The engine adds sentence segmentation, retry of up to 3 attempts per chunk, loudness matching, and silence compression, and the web UI accepts WAV/MP3/OGG/FLAC uploads that are trimmed to 5 seconds, with speed control and manual phoneme editing to fix mispronunciations. The repository has no license file, so the scripts and web interface carry no stated open-source terms. The model weights are CC-BY-NC-4.0, which the README says prohibits commercial use of the output, and an ONNX package is published on HuggingFace as Nimaone/pocket-tts-farsi-v2-onnx.