Why run it yourself
- Data stays with you. What a caller says and hears does not need to leave your building.
- Isolated sites. A site with no internet can still run an assistant.
- Predictable cost. Local models have no per-minute charge from an outside provider.
- Choice. You decide, per assistant, which parts are local and which are not.
The three parts of a voice assistant
| Part | What it does | Local option |
|---|---|---|
| Speech to text | Turns the caller's words into text | Whisper models in several sizes, from a tiny one of about 80 MB to a large one of about 1.6 GB |
| Language model | Decides what to say | Open-weight models served by a container on your own server |
| Text to speech | Speaks the answer | On-server voices, covering 14 languages |
The honest trade-off
Speed is the main cost. On a modest server with only CPUs, a local language model is very good for learning, for text and for background work, but a live conversation is faster on an outside model or on a larger server. The first turn of a call is the slowest, and later turns reuse what was already processed.
A sensible middle path is to mix. Use local speech and voices where privacy matters most, an outside model where speed matters most, and a fallback in either direction if one provider fails.
Choosing sizes and models
- Start from the call you need to support: how many at once, which languages and what answer time you can accept.
- Pick the smallest speech model that is accurate enough on your callers' audio, then measure its speed on your own server.
- Test the language model for a normal reply, a follow-up, memory of the call and a judging role before you rely on it.
- Create knowledge bases with local embeddings, so your documents are not sent out to be indexed.
- Decide fallbacks. If a local part is too slow or down, which outside provider, if any, may take over?
Running with no internet at all
The platform can be installed from an offline bundle on a disk that is verified before it installs. In one test on a virtual machine with the internet blocked, a bundle of about 74 images and 14 GB took roughly 26 minutes to write, 3 minutes to verify and about 40 minutes to install, with no download or build on the new server. That is one test, not a promise for every site. Some features, such as outside AI providers and SMS or email delivery, still need the internet while running, and a fully isolated site uses the local models.
How SSONE does private AI
Speech to text, voices and language models run on the platform's own CPUs in their own container, with five Whisper sizes, 18 on-server voices in 14 languages and open-weight language models with a Test button. Choices are made per assistant, per group or per account, and industry packs ask where their AI should run.