Model Configurations
Model configurations set the models every agent in your organization uses, unless an individual agent overrides them. Open Models in the dashboard:
app.decibyl.ai/model-configurations
Two decisions, and they are independent
The screen asks for one thing at the top and one thing per model, and keeping them apart is the point.
How the conversation is produced is the architecture, and it is the only top-level choice:
| What it is | When it wins | |
|---|---|---|
| Transcriber → Model → Voice | Three models in a chain | More control over each stage, better Indic support, cheaper at most volumes |
| Speech to Speech | One model hears and speaks | Lower latency and more natural interruptions, but no separate transcriber or voice to tune |
Whose key each model runs on is a property of that model, not a mode you enter. Pick Decibyl as a slot’s provider and that slot is managed — we hold the key and bill it through your balance. Pick a vendor and paste your own key and that slot is yours, billed by them.
Finding the model you want to change
The slots are grouped by what they decide. Voice holds the transcriber and the voice — everything that turns speech into text and back. Model holds the language model and the embedding model, which decide what the agent understands and remembers. On Speech to Speech the Voice group collapses to a single slot, because one model is doing both jobs.
Each slot asks for a provider and a model. Everything else it accepts — speed, temperature, the turn-taking timings — is behind Advanced settings, folded away with its default shown beside each field, because most of them are right as they are. A slot you have already changed opens that section for you, so a setting you made is never hidden behind a click.
What the stack costs, before you call anything
As you change a slot the screen re-prices the whole stack per minute. That is the number to look at while choosing — not the vendor’s per-token or per-character price, which is not comparable across a transcriber, a language model and a voice.
A slot with no rate on file reports as unpriced rather than as free, and the estimate says so instead of quietly under-reporting.
Speech to Speech
One realtime model handles speech in and speech out, so there is no separate Voice or Transcriber to configure. You still choose a language model alongside it: variable extraction and QA run on the transcript after the call, and those are text tasks.
Bringing your own keys
Add a provider key once under Provider Keys, then choose that provider on any slot. Keys are encrypted at rest and are never returned by the API after they are saved.
Two things worth knowing:
- Your provider bills you directly. That spend does not appear on the Tokens tab in Decibyl — an empty By model table on an account making calls usually just means BYOK, not a broken meter.
- A key that expires fails at the moment the agent tries to speak, which sounds like silence on the call rather than like an error. If an agent connects and then says nothing, check the key before you check the prompt.
When a managed slot says coming soon
A slot offered as coming soon means the platform holds no key for the provider that tier resolves to. It is not a plan restriction and there is nothing to buy — it turns into a normal choice on its own once the key is in place. Choosing your own provider for that slot works today.
Agent-level model overrides
You can override the organization configuration for one agent — different agents often want different voices, and a Hindi collections agent and an English support agent rarely want the same stack.
- Open the agent.
- Go to Settings.
- Open Model Overrides.
- Enable the override for the service you want to change.
- Configure the provider, model and keys for that service.
- Save.
Overrides are selective: override only the Voice for one agent and it keeps using the organization’s language model and transcriber.