VoxLoom — Self-hosted voice AI studio

Own your voice stack. Run state-of-the-art text-to-speech and voice cloning on your own GPU. No subscriptions. No rate limits. No vendor lock-in. No per-character billing.

The open-source alternative to ElevenLabs that you actually own.

🎧 Listen

Clip	Language	Voice	Generated by
	English	OS default	pyttsx3 (works offline, zero setup)
	English	auto	OmniVoice on cloud GPU
	Chinese	auto	OmniVoice on cloud GPU

🎯 Right now you can listen to row 1 (committed demo/hello_system.wav). Rows 2-3 will appear once a contributor runs scripts/make_demo.py on a cloud GPU (T4 16GB, ~$0.2–0.4/hr) and commits the WAVs. The pyttsx3 track sounds like your OS TTS — neural voices need a GPU.

Why VoxLoom

• 🔓 Open source & self-hosted — MIT. Audit it, fork it, own it.
• 🚫 No rate limits, no quota, no subscription — your GPU, your rules.
• 💸 No per-character billing — commercial TTS charges you forever; VoxLoom is a one-time GPU cost.
• 🌍 600+ languages via OmniVoice (planned in CLI); studio-grade Chinese via Qwen3-TTS (planned).
• 🎙️ Voice cloning in seconds (planned) — 3s reference audio, or text-prompt a voice.
• ☁️ Optional hosted API — when you need more horsepower or a premium voice.

VoxLoom vs commercial TTS

	VoxLoom (self-hosted)	ElevenLabs	Azure TTS
Upfront cost	Free (MIT)	Free tier, then paid	Pay per char
Per-character billing	❌ never	✅ yes	✅ yes
Rate limits	❌ none	✅ yes	✅ yes
Data leaves your machine	❌ no	✅ yes	✅ yes
Voice cloning	planned (3s ref)	✅	✅
Self-hostable	✅	❌	❌

Status (v0.1)

• Bundled & working locally: system engine — pyttsx3 (offline OS speech synthesis, zero model download, zero build, runs everywhere). Smoke-test engine; proves the no-cloud / no-rate-limit pipeline.
• Neural engine interface ready: kokoro (Kokoro-82M, Apache-2.0, EN+ZH) — install on a normal machine or cloud GPU (see note below on the Windows safe-delete build issue).
• Planned (cloud GPU): OmniVoice (600+ langs), Qwen3-TTS (best Chinese cloning), Chatterbox (general cloning), VibeVoice-1.5B (warm-female "Xinran" preset voice).

Install note: on some sandboxed Windows environments, pip building of docopt/jieba (pulled by kokoro/misaki) is blocked by a safe-delete interception. Install neural engines on a normal machine or a cloud GPU instance — the offline system engine needs nothing.

Quick start

pip install pyttsx3
python voxloom.py tts --engine system --text "Hello from my own GPU. No cloud, no rate limit." --out hello.wav

List engines:

python voxloom.py engines

Engines

Engine	Languages	Cloning	License	Min VRAM
system (bundled, v0.1 default)	OS-dependent	no	OS/bundled	0 (CPU)
kokoro (interface ready)	EN + ZH	no	Apache-2.0	~1 GB (CPU ok)
OmniVoice	600+	3s ref / text-prompt	verify LICENSE (k2-fsa)	~6 GB
Qwen3-TTS	10+	yes	Open	8 GB
Chatterbox	23	few-sec ref	MIT	8 GB
VibeVoice-1.5B (Xinran)	ZH	preset voice	MIT*	6–8 GB

* VibeVoice weights are MIT; TTS inference code lives in the community fork vibevoice-community/VibeVoice. OmniVoice license: confirm the LICENSE file in k2-fsa/OmniVoice before any commercial use.

Need more? Hosted API with premium voices

Some voices need bigger GPUs or a commercial SLA. VoxLoom Cloud (planned) offers:

• Standard engines, billed per character (no monthly minimum)
• Premium voices — including our signature warm-female tone, tuned for narration
• Priority queue, higher concurrency, commercial license included

Join the waitlist → https://18aa20b9800242d0959a4125a2858491.bj4.agentos-app.net

License

Repo: MIT. Each engine carries its own license (Kokoro Apache-2.0, OmniVoice verify, etc.) — see docs.