First identify which part has failed
| Symptom | Likely subsystem to check | Confirmed guidance |
|---|---|---|
| Your spoken words never become text or are delayed | Speech recognition: local Whisper / Moonshine / optional Whisper Large Cloud in eligible lobbies | Check Settings > Speech, the phone mic panel and current version; old Deepgram setup no longer exists. |
| You can type but AI caller audio is missing or delayed | Caller voice generation / Qwen3 TTS | Look for the game’s hardware compatibility or fallback warning. |
| Microphone input disappeared after unplugging a headset | Device selection or Windows audio input | Re-select your device in Speech settings; later patches improved reconnection. |
| Some PCs freeze while starting Whisper | Local recognizer / CPU compatibility | Developer notes describe Moonshine fallback and startup timeouts. |
| An old guide asks for a Deepgram API key | Historical v0.1.82 setting removed in v0.1.84 | Do not buy or enter a Deepgram key. Use supported local modes or the new optional Whisper Large Cloud when eligible. |
| The game warns speech recognition cannot run | Hardware compatibility | Use typed caller responses rather than treating the game as entirely unplayable. |
On narrow screens, scroll the table to read every column.
Choose a speech recognition provider
Whisper: a local recognition choice, but some hardware encountered slow starts or compatibility problems. October 18 update 0.1.82 introduced 15-second startup/recognition limits and automatic Moonshine fallback for supported languages if Whisper fails. A saved choice may remain Whisper even when the session temporarily falls back; do not mistake that for evidence the fallback did not happen.
Moonshine: developer notes for 0.1.78 identify English, Arabic, German, Japanese, Mandarin Chinese and Spanish support in the six-language setting. The required Moonshine packs are bundled with the game, rather than an additional download in the menu.
Deepgram is retired: it was an October 0.1.82 historical option, not the current provider. The developer explicitly removed Deepgram in v0.1.84 (October 24); do not follow old instructions to purchase API usage or insert a key into a control that no longer exists. Whisper Large Cloud is a different, later option introduced with v0.1.99 (November 9) for eligible official AI calls in verified Steam lobbies. Selecting it sends your speech to DeepInfra for transcription. It is not the same thing as the removed Deepgram integration, and its availability depends on the call and lobby context. Developer v0.1.84 · v0.1.99.
Qwen3 warning? Keep the two voice systems separate
Qwen3 text-to-speech concerns generated caller voices, while Whisper/Moonshine concern recognizing your spoken words. The developer’s October 18 update says Qwen3 compatibility warnings and setup prompts no longer block hosting or joining. When Qwen3 setup is unsupported or fails, the game should handle its warning without trapping lobby navigation. This does not certify that all GPUs can run local caller voices.
If a caller has no sound, record whether speech input works separately from whether caller output works. That distinction is useful when reporting bugs.
Practical microphone checks
- Open Speech in the game’s settings, and check the current language, selected recognition provider and device.
- On Windows, verify your preferred input device is selected and that desktop apps have microphone permission.
- Read the game warning: a missing Deepgram key, Whisper timeout and Qwen3 incompatibility are different problems.
- On the current build, try the supported Moonshine fallback or the developer’s documented lighter Whisper Base option if local Small cannot start. Where eligible, an optional Whisper Large Cloud choice may appear; never assume it is available in every lobby or language.
- Try a typed caller response to separate microphone recognition from the rest of the game session.
- If the symptom persists, use the game’s feedback/report tools and include OS, CPU, GPU, input device, recognition provider, chosen language and warning text.
Steps 2 and 5 are general reversible diagnostics; no GGClue first-hand test is claimed.
Privacy and recording settings
Update 0.1.78 introduced Disable Call Saving, and the developer states that boss reviews may still use temporary recordings. It also introduced an earlier optional Deepgram provider, removed with 0.1.84. The later v0.1.99 Whisper Large Cloud sends eligible call speech to DeepInfra when selected or used as the documented fallback. Do not infer from the switch name that no audio is ever captured, transmitted or temporarily processed. Read the current in-game policy before choosing cloud transcription, streaming or recording sensitive material.
FAQ
Will Moonshine recognize every language? No. The patch explicitly lists six supported languages in that option.
Do I need to pay for Deepgram? No. Do not buy a Deepgram plan for the current game: the integration was removed in v0.1.84. Optional Whisper Large Cloud from v0.1.99 is a different service, using DeepInfra in eligible official AI calls. Check the current game’s terms and notices rather than assuming the old Deepgram billing instructions apply.
Does a Qwen3 error mean my microphone is broken? No. It refers to caller voice generation rather than microphone recognition.
Related: Game hub · October 18 patch explanation.
November 9–10 current-build verification and success checks
v0.1.99 (November 9): optional Whisper Large Cloud is selectable in Settings > Speech or the phone microphone panel for eligible official AI calls in verified Steam lobbies. It sends call speech to DeepInfra. Moonshine may be absent for languages it does not support; do not treat that as an installation error.
v0.1.100 (November 10): local Whisper Base and Small have lower Windows memory-headroom requirements (1,152 MiB and 1,792 MiB, respectively). On a Small failure the game can retry Base; if local models fail it can release memory and use cloud for eligible calls while preserving the saved preference. This is not proof of a universal speech fix or that cloud calls bypass privacy disclosures. See exact developer patch.
Observable success: after choosing a mode supported for the call and language, a spoken test line appears as recognized caller text; if audio playback fails but typed input works, investigate caller voice output separately. For persistent failure, record the current version, selected provider, call/lobby type, language, available RAM and the actual error.
Sources: historical 0.1.82 patch, v0.1.84 Deepgram removal, v0.1.99 Cloud launch, v0.1.100 Whisper fixes, official Steam listing.
Independent GGClue help page; source rechecked November 11, 2026. No firsthand hardware benchmark, microphone recording, paid service integration, or private data access is claimed.