Does Paradee 1.2 work in the EU?

Available in the EU. Open-weights model reported as Apache 2.0 and run locally via a pip package, so it can be downloaded and used anywhere including the EU; no regional restrictions found. Last verified 14 Nov 2026.

How much does Paradee 1.2 cost?

Open source, open source. Free to download and run locally; Chickadee extension uses local models. Licence: Apache-2.0.

Where was Paradee 1.2 shown?

In 1 creator video by AI Search. Most recently by AI Search on 1 Nov 2026, at 8:46 in "AI finally sleeps, Gemini 4 Carbon, ETH cracked, energy-based models, Anthropic IPO: AI NEWS".

Paradee 1.2

Website GitHub

8.5 MB, 8M-parameter text-to-speech model distilled from Kokoro-82M, with four preset voices and CPU inference.

Available in EUVoice & AudioModel

Updated 14 Nov 2026EU status checked 14 Nov 2026Last shown 1 Nov 2026

Paradee 1.2 at a glance

What is Paradee 1.2?: Paradee 1.2 is a model. 8.5 MB, 8M-parameter text-to-speech model distilled from Kokoro-82M, with four preset voices and CPU inference.
Does Paradee 1.2 work in the EU?: Available in the EU. Open-weights model reported as Apache 2.0 and run locally via a pip package, so it can be downloaded and used anywhere including the EU; no regional restrictions found. Last verified 14 Nov 2026.
How much does Paradee 1.2 cost?: Open source, open source. Free to download and run locally; Chickadee extension uses local models. Licence: Apache-2.0.
Where was Paradee 1.2 shown?: In 1 creator video by AI Search. Most recently by AI Search on 1 Nov 2026, at 8:46 in “AI finally sleeps, Gemini 4 Carbon, ETH cracked, energy-based models, Anthropic IPO: AI NEWS”.

Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.

About

Paradee-8M-v1.2 is a small text-to-speech model distilled from Kokoro-82M. It has about 8 million parameters and fits in an 8.5 MB file. Version 1.2 extends the single-voice v1.1 to four preset voices: af_heart, bf_isabella, am_puck and bm_fable. It does not support voice cloning.

For v1.2, each half of the model (text side and decoder) gets a small table with one voice vector per voice, and the selected voice picks the row. The rest of the model is shared. The model was trained further from the v1.1 weights on 12,000 sentences read by Kokoro in each of the four voices. The phonemizer was updated to use British phonemes for British voices.

The author reports UTMOS scores of 4.40 (af_heart), 4.06 (bf_isabella), 4.32 (am_puck) and 4.18 (bm_fable), compared with 4.52, 4.41 and 4.47 for Kokoro-82M on the voices shown. Whisper-measured WER is in the 5.4-6.3% range. The af_heart score is about the same as in v1.1 (4.40 vs 4.39). Quality for the added voices is below Kokoro's.

It can be used through a Python package (pip install from GitHub, with a command-line --voice option) and as the light voice model in Chickadee, a Chrome extension that reads web pages aloud with local models. Per a YouTube creator summary, it is Apache 2.0 licensed and runs on CPU.

Key features