Three Public Servers for OVOS: Speech, Voice and Translation

TigreGótico Lda hosts three public endpoints for the OpenVoiceOS community, in Portugal:

service	endpoint	engine
Speech to text	stt.openvoiceos.pt	onnx-asr
Text to speech	tts.openvoiceos.pt	phoonnx
Translation and language ID	translate.openvoiceos.pt	linguonnx

Each host serves its interactive API reference at /docs and a health check at /status; the bare root returns 404, which is normal.

Point an OVOS device at them and it speaks, listens and translates without downloading a model. That is the whole idea: a Raspberry Pi or a similar low-power board offloads the heavy part over HTTP and keeps its CPU for everything else.

Live uptime and response times are on the status page.

Read this before you rely on them

These servers are a best-effort community service, offered with no warranty of any kind. Please read this part:

• There is no uptime guarantee and no SLA.
• There is no support commitment. Nobody is on call for them.
• They can change, move or disappear, without notice.
• They are not for production use. Do not put a product, a customer or an automation you care about behind them.
• If you depend on any of this, run your own instance. Every piece is open source and installable, and the servers exist so that you can decide whether the stack is worth self-hosting.

They are here to be tried, demoed, and used on small hardware by people who accept that the endpoint may be gone tomorrow.

What each one is

onnx-asr is a lightweight ASR package that runs modern speech recognition models on onnxruntime, with no PyTorch, no Transformers and no FFmpeg. The public endpoint chooses a model per language, so a client passes audio and a lang and gets back a transcript.

phoonnx does multilingual phonemization and text to speech over ONNX voices, reaching more than a thousand languages and voices across its bundled voice indexes. The endpoint holds one default voice per supported language, and phoonnx.openvoiceos.pt takes a voice name per request when you want a specific one.

linguonnx does machine translation and text-side language identification on the CPU, again through onnxruntime and without torch. Its registry reaches 593 languages over the default model graph, and when no single model covers a pair it chains small models through a pivot language. The route depends on the limits the caller sets (a size budget or a hop preference can send the same pair through a different pivot, or through one large multilingual model in a single hop), and the translator reports which models and pivots it used: the README's own example translates Portuguese to English through opus-mt-pt-en-int8, a 172 MB model, and route.pivots names any intermediate language. The routing detail is in the linguonnx README.

A translation is one GET:

$ curl 'https://translate.openvoiceos.pt/translate/pt/hello%20world'
"Olá mundo"

Pointing OVOS at them

{
  "stt": {
    "module": "ovos-stt-plugin-server",
    "ovos-stt-plugin-server": {"url": "https://stt.openvoiceos.pt/stt"}
  },
  "tts": {
    "module": "ovos-tts-plugin-server",
    "ovos-tts-plugin-server": {"host": "https://tts.openvoiceos.pt"}
  },
  "language": {
    "detection_module": "ovos-lang-detector-plugin-server",
    "translation_module": "ovos-translate-plugin-server",
    "ovos-lang-detector-plugin-server": {"host": "https://translate.openvoiceos.pt"},
    "ovos-translate-plugin-server": {"host": "https://translate.openvoiceos.pt"}
  }
}

The client plugins are thin HTTP wrappers: ovos-stt-plugin-server, ovos-tts-plugin-server and, for both translation and language detection, the package ovos-translate-server-plugin, which provides the two *-plugin-server entry points in the block above. They hold no model and no engine, so a device that uses them installs almost nothing.

The servers also carry the vendor-compatible routers of their packages (OpenAI, ElevenLabs and others), but on this host the OpenAI speech route POST /openai/v1/audio/speech returns 500 while the native /synthesize route answers, so the compatible routes are not something to rely on here.

What to expect, honestly

Warm requests are quick. Cold ones in rare languages are not. A model that is already loaded answers fast. A language served only by a large multilingual model has to load that model first; the registry's language-identification models alone range from 33 MB to 1.7 GB. That first request can take minutes. Ask again and it is fast, until the model is evicted to make room for another.

Routable is not the same as usable. 593 languages route somewhere, and a route is only as good as the models on it; the registry records per language what each model covers, and the servers inherit it. Treat the tail as something to test before you trust it, not as a supported language list.

Shared machines are shared. These endpoints have no per-user quota and no queue guarantees. Somebody else's cold multi-gigabyte load is your slow request.

Run your own

Every service here is a package you can install:

• onnx-asr behind ovos-stt-http-server
• phoonnx behind ovos-tts-server
• ovos-plugin-linguonnx behind ovos-translate-server, which turns any OVOS language plugin into a micro service

The same client configuration above works against your own host: change the URLs. That is the point of the public trio: try the stack cheaply, then own it. Problems with the endpoints themselves are best reported on the engine repository concerned; there is no tracker for the hosts.

---

This work is part of the OpenVoiceOS From Beta to Breakthrough milestone, funded through the NGI0 Commons Fund, a fund established by NLnet with financial support from the European Commission's Next Generation Internet programme, under the aegis of DG Communications Networks, Content and Technology under grant agreement No 101135429. Additional funding is made available by the Swiss State Secretariat for Education, Research and Innovation (SERI).

---

Help Us Build Voice for Everyone

OpenVoiceOS is more than software, it's a mission. If you believe voice assistants should be open, inclusive, and user-controlled, here's how you can help:

• 💸 Donate: Help us fund development, infrastructure, and legal protection.
• 📣 Contribute Open Data: Share voice samples and transcriptions under open licenses.
• 🌍 Translate: Help make OVOS accessible in every language.

We're not building this for profit. We're building it for people. With your support, we can keep voice tech transparent, private, and community-owned.

👉 Support the project here