First Officer and Small Talk

English | Deutsch

First Officer & Small Talk

First officer

With the voice loop on, every flight gets its own first officer:

• a name picked at random from internationally easy-to-pronounce names (e.g. Anna, Emma, Sophie / Sam, Ben, Daniel),
• a voice of the same gender that is never the voice of an ATC station.

Name and voice stay for the whole flight – also across Duopilo restarts (%LOCALAPPDATA%\Duopilo\first-officer.json). A new SimBrief plan means a new flight and a new first officer. The voice is used for callouts, replies and small talk.

Shown	in the dashboard header ("F/O Sam"), the voice panel and Settings
Introduction	once per flight at the gate (on the ground, within 5 NM of the departure airport) – "Hi, I'm Sam, I'll be your first officer today" – or when you say "Briefing"
Addressing	"Sam, gear up" works like "Kopilot, gear up". Whisper's spelling variants are tolerated (Erik/Eric, Jonah/Jonas), ordinary words never count as the name ("nice" isn't Nico)
Fixed name	Settings → First officer and small talk (with gender for names not on the list)
Off	same place – Duopilo then speaks with the dashboard voice

Voice

Settings → First officer and small talk → Voice: Automatic (per flight, matching the name) or a fixed voice, with Listen to preview. Saved as voices.kopilot in voice-settings.json, applies immediately. A fixed voice decides the gender.

Engine	Voices	Speaks
Kokoro	female af_heart, af_bella, bf_emma, af_aoede, af_kore, af_sarah; male am_michael, am_fenrir, am_puck (reserve bm_george, bm_fable)	English only – small talk is then English too
Supertonic 3	M1, M2 (male), F1 (female)	callouts in English, small talk in German

Automatic picks Supertonic when small talk is German, otherwise Kokoro. Until the Supertonic model (~400 MB, SHA-256-checked) is downloaded, the first officer speaks with a Kokoro voice. Don't want Supertonic: Smalltalk:Supertonic: false.

Callouts and small talk always come from the same speaker.

Small talk (optional, off by default)

A sentence that is neither a command, a query nor a radio call goes to a language model instead of "Say again":

"Hey Sam, how's the weather in Berlin?"

Provider	What you need
Claude	your own API key (console.anthropic.com); model empty = default Haiku model
OpenAI-compatible	base URL (default https://api.openai.com/v1; also Groq, LM Studio, vLLM …), model, API key if required, Reasoning
Ollama	address of another computer (e.g. http://192.168.1.20:11434) and the model

There is deliberately no local model inside Duopilo – it would cost too much GPU and RAM next to MSFS. Run Ollama on a second machine if you want it local.

Set up under Settings → First officer and small talk (applies immediately; Test asks a short question). The API key is stored only in appsettings.Local.json (or Duopilo__Smalltalk__ApiKey), never sent back to the browser and never written to logs or reports.

Rules

• Read-only. The model gets read-only context (flight phase, position, route, next waypoint, ETA, fuel and fuel against plan, wind and temperature, METARs of the plan, ATC station) and the last few sentences. Its answer is only spoken – never executed as a command and never prepared for "Confirm".
• Lowest priority. Callouts, warnings, acknowledgments, checklists and radio always come first; anything interrupts small talk (no repeat). Nothing goes to the model while a readback waits, a checklist runs, a message waits for "Check" or the mic is muted.
• No room chatter. Sentences without addressing go to the model only up to UnaddressedMaxWords words (default 12).
• Limits. MaxReplyWords (40), RequestsPerMinute (6), TimeoutSec (10). Above the limit, offline or on any error: "I can't chat right now".
• Language. Small talk follows the dashboard language (if the voice can speak it); callouts and radio stay English.

Reasoning models

Reasoning models (e.g. openai/gpt-oss-20b on Groq) spend tokens thinking, which can leave nothing for the answer. Reasoning (OpenAI-compatible only, Smalltalk:Reasoning):

Setting	Behaviour
Automatic (default)	sends reasoning_effort: "low" for models with "gpt-oss" in the name, nothing otherwise
Off	never sends it
Low / Medium / High	always sends it; adds token headroom (+512 / +1024 / +2048)

An empty answer at the token limit gets its own warning in the log and on the Test button.

Privacy and costs

The settings page states where the conversation goes. With Claude or a cloud OpenAI-compatible service, your sentences and the read-only flight context are sent to that provider and billed to your key.

Settings reference

"Duopilo": {
  "FirstOfficer": { "Enabled": true, "Name": "", "Gender": "auto", "Introduce": true },
  "Smalltalk": {
    "Provider": "off",          // off | claude | openai | ollama
    "Model": "", "Url": "",
    "MaxReplyWords": 40, "RequestsPerMinute": 6, "TimeoutSec": 10, "UnaddressedMaxWords": 12,
    "Supertonic": true, "SupertonicThreads": 2,
    "Reasoning": "auto"
  }
}