Offline AI Voice Chat on Windows: Setup and Mic Fixes

Offline AI voice chat on Windows needs a working microphone and three local stages: speech recognition, chat generation and speech output. Install their files while online, check each stage separately, then repeat after restarting without a network. A moving microphone meter alone does not prove that the conversation is offline.

For Local Waifu, begin with a supplied voice rather than a custom clone. This procedure reviews the Windows implementation and Microsoft’s guidance as of October 22, 2026. It is a source-based walkthrough, not a Windows runtime test or a speed benchmark.

Prepare all the local parts before disconnecting

The short version: Download the chat and speech assets, then explicitly choose local processing.

Check the Windows hardware requirements before selecting a chat model. Keep the first test simple: one character, a short English sentence and no online tools.

1. Install and select a local chat model under Settings > Models. A saved cloud model is not an offline substitute.
2. Open Settings > Models > Voice Model. Expand Models and files near the bottom if needed.
3. Install Speech recognition and the Voice model. When both are missing, Set up voice installs the two in sequence. Wait for completion, not just the start of a progress bar.
4. Turn off Use cloud speech recognition and Use a cloud voice. These are separate routes: one can send recorded audio away; the other can send reply text away for speech generation.
5. Under Settings > Performance, enable Offline Mode in the network section. It blocks supported cloud AI routes, including cloud speech. It is not a firewall for updater or licensing traffic.

The Windows recognition path uses a local whisper-server.exe helper and downloaded Whisper weights. Its server binds to 127.0.0.1, meaning this computer. The implementation can select a bundled CPU helper when a fully provisioned CUDA helper is absent; do not begin mic troubleshooting by forcing GPU settings.

Start with the supplied Supertonic voice. Natural voice, used for custom voice references, is a separate Chatterbox runtime and model. Windows code gates enabling it on at least 16 GB RAM. Installing that runtime is not the same as having its model loaded and ready: model files can download on first startup. Keep this optional route off until the basic test passes.

The voice setup guide maps the controls. For background on the separate stages, see how offline voice works; do not treat that article’s Mac timings as Windows expectations.

Get a clear signal in Windows first

The short version: Windows must detect the intended input before speech recognition can help.

Check the headset’s physical mute switch and connection. For Bluetooth, confirm that it is paired and connected. Microsoft also recommends selecting the microphone as the system default recording device.[2]

On Windows 11, open Settings > Privacy & security > Microphone. Enable Microphone access, Let apps access your microphone, and Let desktop apps access your microphone. Desktop apps do not use the same individual switches as Store apps, so a missing Local Waifu entry there is not itself a failure.[1][2]

On Windows 10, use Settings > Privacy > Microphone. Check device microphone access, Allow apps to access your microphone, and Allow desktop apps to access your microphone.[1] These paths are for existing installations; Microsoft lists Windows 10’s standard support as ended.[1]

Next, open Settings > System > Sound > Input and select the intended microphone. In Windows 11, open its properties and use Start test, speak normally, then Stop test and Play. Adjust Input volume if the sample is too quiet.[3] In Windows 10, speak while watching Test your microphone for movement.[2]

Pass: Windows shows input activity, and the Windows 11 recorded sample contains your voice. Fail: fix the connection, mute state or selected input before changing an AI model. If that does not help, use Microsoft’s input-audio troubleshooter.[2] Driver removal is not part of this first-pass procedure.

Check Local Waifu’s input without starting a conversation

The short version: Match the app’s input to the Windows device, then stop its meter test before dictating.

In Settings > Models > Voice Model, find Microphone and speaker. The reviewed Windows interface has Microphone and Speaker pickers, plus System default. Calls and dictation use the saved microphone choice.

Start with System default after choosing the correct Windows input, or select a listed device explicitly. Choose Test microphone and speak: the level segments should move. Select Stop afterward. Dictation refuses to start while this microphone test is active.

After unplugging a headset, check the picker again. The implementation can fall back to Windows’ default if a saved device name disappears. Reopen the settings panel after connecting a different device to refresh its list.

Now use the chat composer’s microphone button, Dictate. Say a sentence, then select Stop dictating. Expect editable text in the composer, not an automatically sent message. This is a useful speech-recognition test because it does not depend on generating or hearing an answer.

Follow the first failed stage, not the loudest symptom

The short version: Fix the earliest failing check and leave the working stages alone.

Under Settings > Performance > Runtime status > Check, inspect Local AI engine, Chat model and Voice output. Treat Ready as a useful clue, not an end-to-end audio test: the Windows voice readout can recognize installed voice files without proving audible playback.

Do not rely on an older fallback voice if the local voice model is missing. The reviewed Windows build does not provide the usable legacy TTS sidecar that a Mac-oriented description might imply. Install the supported voice assets and test them. Keep I’m wearing headphones off when using speakers; the interruption option can otherwise let her own output trigger the microphone.

For errors that survive these checks, use troubleshooting and the hardware and network guide. Note the exact error before changing another setting.

Run a repeatable offline conversation test

The short version: Repeat the same short chain after a cold start with every network connection unavailable.

This is an example to perform, not a reported test result. Use headphones and keep the supplied local voice selected.

1. Windows input: confirm the chosen microphone shows activity; play its recorded test sample where available.
2. Transcript: dictate “I would like a quiet evening at home.” Stop recording and check that the composer contains those words. Do not send it yet.
3. Local text reply: clear the composer, type “Suggest one quiet activity for tonight,” and send it with the local chat model selected and Offline Mode on. Expect a new text answer.
4. Speech output: end any call, open Test her voice, type “A quiet evening sounds good,” and choose Speak. Expect audible speech. This preview uses webview audio playback, so also verify the Speaker selected for the real call.
5. Combined call: start with the phone button, wait until it is listening, say the first sentence and pause. Expect recognition, a fresh reply and spoken output. End the call.
6. Cold offline repeat: close Local Waifu, disconnect Wi-Fi and Ethernet, and disconnect any other active network path. Restart the app and repeat these checks, using a different activity request so you can distinguish a new answer from old chat history.

If only the cold offline run fails, record which stage broke. Missing downloads and an accidentally selected cloud route need different fixes. Reconnect only to finish the identified setup step, then repeat from a disconnected restart.

A passing run supports the offline capability of that configuration. It does not establish that every background process makes no connection attempts. The offline companion setup overview explains that wider boundary. Keep a short record of your app version, selected models, microphone and first failing stage; that is more useful for support than “voice is broken.”

Sources

The short version: Microsoft supports the Windows checks; product steps come from Windows source inspection, not a live Windows test.

[1] Turn on app permissions for your microphone in Windows

[2] Fix microphone problems

[3] How to set up and test microphones in Windows

Questions people ask

Back to the blog