Shape OVS evolves more than ElevenLabs for your local voice ai 💡

[debpalash]

Help shape how OVS becomes a true local ElevenLabs alternative

Hey everyone 👋

First, a personal note: much of APAC — India especially — is suffering badly under the super/El Niño heatwaves right now, and people are dying, many of them elderly. Climate change needs global effort, not just local. I'm using my voice for awareness here — please come together and do something where you can.

Back to OVS: ~20 PRs landed today fixing bugs you reported here and on GitHub. Thank you for the reports — keep them coming.

The goal for the coming months: OVS competing with ElevenLabs on feature parity, quality, and freedom — without subscriptions, accounts, or your voice ever leaving your machine. Here's what's on the table, now fully researched in docs/competitive-analysis.md (licenses, prior art, regulations, effort sizes):

• Agentic voice workflow — run OVS on your desktop/server and let it act as your voice: an assistant that speaks for you in agent pipelines (works with pipecat/LiveKit today via the OpenAI-compatible API), a persona bot for your communities (like omnivoice_bot on this server), and eventually opt-in phone calls — with consent locks, disclosure, and watermarking built in from day one, not bolted on.
• Production audiobook/stories creator — ebook in, chapterized m4b out, mastered to ACX technical spec, editable as a multi-track project.
• Persona Gallery — share and import voice personas with consent attestation and watermarked previews; a portable .ovsvoice format so your voices aren't locked in.
• Remote GPU backend — run inference on your GPU box, drive it from a thin desktop client over your tailnet, with proper auth (unlike the 175k exposed Ollama servers out there).
• Unlimited-length generation — chunked + crossfaded, no more length ceiling.
• WisprFlow-grade dictation that complements agentic/code editors — local LLM cleanup of filler words and Whisper artifacts before paste.
• Remote Ollama/vLLM/OpenAI-compatible endpoints configurable right in the app UI.
• Polished dubbing experience — dub-length fitting, per-segment clone refs, second-pass timing QC.
• Better model & env management + a real GPU compat matrix (MLX / NVIDIA / AMD / CPU) — so "works on my 5090/1080" stops being a guessing game, with clear preflight errors instead of silent CPU fallbacks.
• Polished OpenAPI docs with Scalar — already live at /docs, spec hygiene incoming.

Want to influence priorities or PR something yourself? The competitive analysis has an implementation spec for every item above. Drop your ideas below 👇 🎙️

[richardchen874-sys]

The local ElevenLabs-alternative roadmap should still measure voice workflow cost per completed audio task, especially when OpenAI-compatible endpoints sit behind Pipecat or LiveKit.

I would define a small provider capability matrix and attach usage/latency/error metadata to every routed call. That gives you a clean path to cost-aware routing without hiding why a route was chosen.

This is close to what I am experimenting with: official Chinese models behind an OpenAI-compatible multi-model layer, with an emphasis on predictable usage and routing behavior. For OmniVoice-Studio, would you rather optimize routing first for cost, latency, or fewer failed agent steps?