The reject pile

Curating means saying no.

“Hand-picked” is a word. Anyone can type it. There is exactly one proof a dump list can never show: what didn’t make it — and why. Here it is. 6 noes, each with a reason you can check and the date we looked at the source.

• canopyai/Orpheus-TTS

  TTS model · rejected

  Promises

  State-of-the-art open-source speech synthesis with zero-shot cloning, ~200 ms latency. Repo licence Apache-2.0, and tagged apache-2.0 on HuggingFace too.

  Why not

  The code is Apache-2.0. The weights are not. The HF tags identify the model as a finetune of meta-llama/Llama-3.2-3B-Instruct — and Llama weights cannot be relicensed to Apache-2.0. What stays is the duty to display “Built with Llama”, to pass the agreement on, and to “include 'Llama' at the beginning of any such AI model name”. Orpheus does none of that.

  Code licence ≠ weights licence. A licence tag is a claim by whoever uploaded it, not proof.

  checked 31 Jul 2026 · HuggingFace model API + llama-models/models/llama3_2/LICENSE

• coqui-ai/TTS · ⓍTTS-v2

  TTS library · rejected

  Promises

  Voice cloning in 16 languages. The most-linked “clone your voice locally” tutorial on the internet.

  Why not

  The weights sit under the Coqui Public Model License: non-commercial — and that restriction explicitly reaches the output and indirect revenue. A video using this voice that leads to a paid product is exactly that case. Second problem: Coqui as a company no longer exists.

  “Non-commercial” doesn't stop at the model. It can reach into what you produce with it.

  checked 31 Jul 2026 · Repo LICENSE + model card

• SWivid/F5-TTS

  TTS model · rejected

  Promises

  Fast cloning with good quality. Sits in practically every recommendation list next to ⓍTTS.

  Why not

  Code MIT, weights CC-BY-NC-4.0. The repo says so itself: “The pre-trained models are licensed under the CC-BY-NC license due to the training data Emilia.” For a monetised product that is a disqualifier, not fine print.

  Here the README tells the truth — you just have to read past the MIT badge.

  checked 31 Jul 2026 · Repo README + LICENSE

• ElevenLabs · free tier

  TTS service · rejected

  Promises

  The best voice on the market, 10,000 characters a month for free.

  Why not

  Terms of Use, verbatim: “if you access or use our Services free of charge … you may only use the Services for non-commercial purposes.” A monetised channel is commercial. The “Developer API” page changes nothing — same account tiers, no separate developer agreement. And even if it did: 10,000 characters don't cover a real monthly need of ~13,000.

  With hosted services the restriction isn't in the licence, it's in the terms.

  checked 31 Jul 2026 · elevenlabs.io/terms-of-use, read live

• rany2/edge-tts

  TTS client · rejected

  Promises

  Microsoft's neural voices, free, no API key. The exact voice we used ourselves for two months.

  Why not

  It's an unofficial wrapper around the Edge browser's read-aloud endpoint. LGPLv3 covers the client, not the audio it produces — for the output there is no grant of rights at all, neither permitting nor forbidding. No contract, no SLA, switchable off at any time. That's not a litigation risk, it's an operational one: the brand voice hangs off an undocumented endpoint.

  We list it even though we used it ourselves. That is precisely the point of this page.

  checked 31 Jul 2026 · LICENSE file in the repo

• Piper · voices “ryan” and “hfc_female”

  TTS voices · rejected

  Promises

  The two most recommended English Piper voices. Blazing fast, run on a Raspberry Pi.

  Why not

  Piper itself is MIT and perfectly fine. But these two voices in particular are licensed non-commercially on their own. Follow the internet's default recommendation and you land straight in the one corner you can't use.

  With Piper every voice carries its own licence. The project's licence says nothing about the speaker's.

  checked 31 Jul 2026 · Voice metadata in the Piper catalogue

• k2-fsa/OmniVoice

  TTS model · on hold

  Promises

  Zero-shot TTS for 600+ languages, cloning, voice design via attributes. Apache-2.0 on code and weights, verified clean.

  Why not

  Not a no — just not chosen. We publish in English; 600 languages are an advantage we don't cash in, and every extra model in the stack is maintenance. Stays the candidate the moment on-device speech is needed in an app — the same group builds sherpa-onnx.

  Not every no is a flaw. Sometimes it's just “good, but not for us” — and that belongs here just as much.

  checked 31 Jul 2026 · HuggingFace model API + repo LICENSE