Paradee is a small English text-to-speech model with 8.07M parameters, distilled from Kokoro-82M and speaking a single voice, Kokoro's af_heart. The int8 build ships as one 9 MB ONNX file, compared with 325 MB for Kokoro, and the README reports about 18x faster than real time in onnxruntime on one CPU thread with no GPU. On the author's 200 held-out sentences the model scores UTMOS 4.41 against 4.52 for Kokoro, with the same 5.7% word error rate measured with Whisper base in the paper; the released ONNX file is reported at UTMOS 4.41 and WER 6.0% under a slightly different text normalization. The README also compares it with Kokoro-7M-Distill, Piper (en_US-lessac-medium) and KittenTTS nano 0.8 on the same sentences. Training is done in stages against the frozen teacher: Kokoro reads 12,000 WikiText-103 sentences (23.9 hours of audio) with every intermediate value saved, a 4.23M-parameter text side learns durations, pitch, loudness and phoneme features, and a 3.85M-parameter decoder learns to turn those values into audio, first with spectrogram losses and then adversarial training. A parameter-free filter corrects the phase of voiced sound between 2 and 8 kHz to remove a slight buzz. Text is phonemized with misaki, Kokoro's own grapheme-to-phoneme library, and the ONNX graph maps phoneme ids (at most 512 including pad tokens) and a speed value to a 24 kHz waveform. A web/misaki.js helper lets browser phonemizers match misaki's spelling, and Paradee serves as the light voice in the Chickadee Chrome extension. Install with pip from the GitHub repo and run python -m paradee on a sentence to get a WAV file. The code is Apache 2.0, the same as Kokoro, and the training sentences come from WikiText-103 under CC BY-SA 3.0. Model files and audio samples are on the Hugging Face Hub.