Speech Output

English | Deutsch

Speech Output

Duopilo speaks callouts, replies and ATC itself – in English, like a real crew. Claude.ai and voice mode can't receive push messages from a server, so this is how warnings reach you without asking.

Priority

Everything runs through one queue:

Warnings > Callouts > Replies > ATC > Small talk

Warnings interrupt everything; older alerts expire after 8 s. An interrupted reply or ATC message is repeated once after the warning.

Only the tray instance speaks – not the stdio instance Claude Desktop starts, otherwise everything would be heard twice.

Engines and voices

	Windows voice	Kokoro	Supertonic 3
Sound	Windows text-to-speech	neural, natural	neural, natural
Languages	English voice needed	English	English and German with every voice
Download	– (install an English Windows voice)	model ~310 MB on first use	model ~400 MB when the first officer gets a Supertonic voice
Used for	callouts	callouts, ATC stations, first officer	first officer when small talk is German
CPU	practically none	2 threads (Speech:KokoroThreads), fixed callouts pre-computed	2 threads (Smalltalk:SupertonicThreads), ~1.3 s per short sentence

Automatic (default) uses the Windows voice if an English one is installed, otherwise Kokoro. While Kokoro is loading, the browser reads aloud in the dashboard.

With the first officer on, the first officer's voice speaks callouts, replies and small talk – one speaker for everything. ATC stations always get their own Kokoro voices.

Installing an English Windows voice (optional)

Settings → Time & language → Language & region → Add a language → English (United Kingdom) with "Text-to-speech". Restart Duopilo.

Settings

Dashboard → Callouts → Voice:

• speech output: Automatic, Windows or Kokoro
• Windows voice (e.g. George, Hazel) or Kokoro voice (e.g. George, Emma, Michael)
• output device – e.g. the headset instead of the speakers
• volume and rate
• minimum level: all, caution and above, warnings only
• add radio altitudes / minimum / retard (for aircraft without their own)

The panel also shows which voice is speaking and how long Kokoro took for the last callout. Tray menu → Speak callouts switches speech on and off.

Stored in %LOCALAPPDATA%\Duopilo\speech.json; models in %LOCALAPPDATA%\Duopilo\models.

Pronunciation

ATC, callouts, fuel checks and voice-loop replies share one pronunciation (Speech/Pronunciation.cs):

What	Spoken
Five-letter fixes	as a word: HENDY → "Hendy"
Navaids (up to three letters)	name from the OFP navlog: YWK → "Wabush"; without a name in the NATO alphabet: FFM → "foxtrot foxtrot mike"
ICAO codes	always spelled: KLAX → "kilo lima alpha x-ray"
Coordinate points	53N050W → "fifty three north fifty west"
Runways	25R → "two five right", 09 → "zero niner"
Squawks, headings, speeds, frequencies	digit by digit, 9 as "niner", frequencies with "decimal"
Flight levels	"flight level three four zero", whole hundreds "flight level one hundred"
Altitudes	in thousands: "one zero thousand", "two thousand five hundred"
ECAM titles	spelled out: "ENG 1 FIRE" → "engine one fire", "APU" → "A P U"

Speeds are always in knots, never Mach.

Muting

Goal	How
Mute one callout (still shown)	"mutedRules": [ "PM_GEAR" ] in speech.json
Turn a rule off completely	Duopilo:Monitoring:DisabledRules – see Pilot Monitoring
All speech off	tray menu → Speak callouts

Licenses

Kokoro-82M and voices: Apache 2.0 · KokoroSharp: MIT · Supertonic 3 weights: BigScience OpenRAIL-M (use restrictions apply, e.g. disclose machine-generated voice in published content) · Supertonic inference example: MIT.