An open directory of open-source text-to-speech models. Every voice reads the same three scripts so you can actually compare them.
ElevenLabs is expensive. Open-source TTS has caught up — but the models are scattered across GitHub, Hugging Face, and a dozen comparison blog posts. Finding one and actually hearing it usually means cloning a repo, fighting CUDA, and praying the weights still download.
OpenSpeech fixes that:
- Every voice reads the same three scripts.
- Specs, license, and install in one place.
- Filter by license, VRAM, language, capability.
- Add a new model with one PR.
- 22 curated models from the broader OSS TTS ecosystem
- ~40 voices across them
- 3 standardized scripts per voice: neutral, emotional, numbers
- Side-by-side audio so you can pick the model that fits your use case
- Next.js 16 (App Router, static site)
- Tailwind CSS v4
- Replicate for sample generation
- Data lives in plain JSON in
data/— no database
git clone https://github.com/sudomichael/openspeech
cd openspeech
npm install
npm run devThe samples are committed to public/samples/ — you don't need a Replicate token just to browse.
To add a new model or regenerate existing ones:
echo "REPLICATE_API_TOKEN=r8_..." > .env.local
npx tsx scripts/generate-replicate.ts # full run
npx tsx scripts/generate-replicate.ts --missing # only missing samplesSee CONTRIBUTING.md for the full workflow.
We want this to be the obvious place to compare open-source TTS. Adding a model is one PR.
- Model list curated from awesome-ai-voice
- Sample generation via Replicate
- Not affiliated with any model author
MIT for the code. Each model retains its own license — see the individual model cards.