Configure AI services

Silo is in active development before 1.0. These guides are updated frequently.

Silo uses a text model to translate subtitles and descriptions, and a speech-to-text model to create subtitles from audio. You can set up either one on its own. Each feature also has its own switch; once it is on, its actions appear in the apps that support it.

Before using a hosted model, check its charges and data policy. Translation sends text to the model’s endpoint; transcription sends audio.

Connect a text model

1. Open Admin > Integrations > AI Services.
2. Under Providers, choose Add provider, pick the provider’s type, and choose Continue. Enter its API key, or leave it empty for a model on your network that needs none, then choose Add provider. For OpenAI-compatible, also enter its Base URL and choose what it Serves.
3. Under Models, choose Set up on Text model. Set Provider to that provider, or to ChatGPT plan as described in Use your ChatGPT plan, and choose a Model.
4. Under Advanced, choose a Reasoning level. See Choose a model and reasoning level.
5. Choose Test text model. The test uses the values on screen, even before you save.
6. Turn on Translate subtitles, Translate descriptions, or both.
7. Save, then translate one short item before using it across a library.

Description translation for viewers controls translation while people browse: Off, Translate button on detail pages, or Automatic on view. Besides detail pages, it covers the episode list on a season or series page and the Featured banner on Home and on a library’s Recommended tab. With Off, browsing never starts a translation job, and you can still translate from the metadata editor or for a whole library.

Home and library rows show descriptions in the viewer’s language once a translation exists. An AI translation is marked Translated by AI in the web app and the Apple and Android apps. Descriptions from a provider or written by hand have no mark.

Connect speech-to-text

1. Under Models, choose Set up on Speech-to-text. Set Provider to one that serves speech, adding it under Providers first if needed, and choose a Model.
2. Choose Test speech-to-text. The endpoint must return timed segments, which Silo turns into subtitles.
3. Turn on Create subtitles from audio and save.
4. Try one short item in the web player. When the job finishes, select the new subtitle track and check its language and timing.

A ChatGPT plan doesn’t include speech-to-text, so transcription needs its own provider even when translation uses the plan.

Use your ChatGPT plan

Instead of an API key, the text model can use a ChatGPT subscription. Translations then count toward that plan’s usage limits, with no separate per-request charge.

1. Under Models, choose Set up on Text model, set Provider to ChatGPT plan, and choose Continue with ChatGPT.
2. Choose Open ChatGPT sign-in and allow Silo to use your plan.
3. After you approve, the ChatGPT tab moves to a local address and may show a “can’t connect” error, which is expected. Copy the full URL from that tab’s address bar.
4. Back in Silo, paste it into Browser URL and choose Connect ChatGPT. If ten minutes have passed since step 1, start again with Continue with ChatGPT.

Under Model, choose one of the models the account offers, then continue from the Reasoning level step in Connect a text model.

To add another account, choose Continue with ChatGPT again, then pick the one Silo uses under ChatGPT account. To sign an account in again after its sign-in expires, use the Reconnect button that shows its email address. Disconnect signs the active account out and revokes Silo’s access in ChatGPT. If Silo shows Disconnected locally, open Manage usage and check that Silo’s access is gone.

When an account reaches its plan’s usage limit, Silo pauses translation and shows ChatGPT plan usage is paused. Check your usage with Manage usage, then choose Test text model; a successful test resumes translation. An account marked plan sharing unavailable is on a plan that does not let Silo use it, so choose another account or set Provider to one with an API key. While the active account is paused or needs sign-in, subtitle and description translation are unavailable, including for viewers. Transcription keeps working if speech-to-text is set up.

Choose a model and reasoning level

Translation is a short task repeated many times: by default Silo sends a film’s subtitles 40 lines at a time, so a feature film takes about 30 requests. A fast, low-cost model usually handles it well, and a larger model mostly adds time and cost. After you translate your first full film, check its cost on your provider’s usage page before translating a library.

Before you rely on a model, translate a few items into each language your viewers use and read them. Weaker models can drift into a closely related language, such as Danish in a Norwegian translation, or leave names in the Latin alphabet inside Arabic or Chinese text.

Reasoning level sets how much the model thinks before it answers. For models Silo recognizes, it lists only the levels that model supports. Higher levels take longer and use more tokens, which hosted providers bill as output. For translation, start with None or Low, and raise it only if translations in your languages miss the meaning. Provider default lets the model decide, and some models then think on most requests, which makes each one slower.

A model you run on your own hardware costs nothing per request, but the open models that fit in 16 GB of graphics memory translate noticeably worse than low-cost hosted models. Many of them think by default; set Reasoning level to None, or a single request can run for minutes.

Limit cost and server load

Under Server-wide tuning, set Jobs running at once; changing it needs a server restart. Under Per-account limits, choose a transcription allowance and how often it resets.

All profiles in an account share its transcription allowance. An administrator using the account’s primary profile is exempt; the account’s other profiles still count against it.

Each account can start only a limited number of description translations a minute while browsing, and only into a language its profile reads: the profile’s metadata language, or the library’s when the profile doesn’t set one.

Keep the default batch and audio-request sizes unless the provider reports a request-size or rate-limit error.

Generated subtitle tracks are saved on the server, and anyone who can play the item can select them. Manage saved files under Subtitle Files.

A test or job fails

Check the model name, base URL, credential, and the provider’s response. A full film can still fail after a passing connection test, for example when it reaches the provider’s quota. Read generated subtitles before relying on them.