First-time setup

The first time you open ASIST, you go through nine screens in order. Screens that your way of talking does not need are skipped. At the bottom of each screen, one line tells you what to do next.

1. Choose a language

   Choose one of 11 languages. Your system language is selected at first. The language you choose becomes the starting value for three things: the language of the interface, the language of the conversation, and the region. The screens that follow appear in that language. You can change each of the three separately in the settings later. The text that describes ASIST’s character, under Character in the settings, is also in the conversation language: in Japanese for Japanese, and in English for every other language. Until you edit it, it also follows any later change of the conversation language.

   

2. Read what to know before you use it

   Four points appear: answers can be wrong, changes happen only after you approve them, each provider bills you for API use, and the conversation goes to the providers you chose. Once you have read them, tick “I have read and understood these four points” to go on. Using ASIST safely explains them in full. If you finished the setup before this screen existed, the same points appear once the next time you open ASIST.

   

3. Enter an API key for the conversation model

   Choose a provider, enter your API key and click “Verify and save”. ASIST checks the key by actually sending a request, and saves it only when it works. The conversation uses the provider’s standard model, and the bridge phrase uses a light, fast model. You can change both in the settings later.

   If you choose “ChatGPT”, “Continue with ChatGPT” appears instead of the key field. Once you sign in with ChatGPT in the browser, the conversation runs on your Plus or Pro plan. See Using your ChatGPT plan.

   

4. Choose how you talk

   Choose “Talk by voice”, “Talk in real time with a voice model”, “Type, and hear the replies” or “Text only”.

   With “Talk in real time with a voice model”, Gemini Live listens and speaks, so the screens for speech recognition and reading aloud are skipped. Gemini Live needs a Google API key. If you verified a Google key on the previous screen, you do not enter it again. A key saved earlier or set in an environment variable is verified when you choose this option, so you do not enter it again either. If there is no key, enter it on this screen and click “Verify and save”. The voice is Google’s default, and you can change it in the settings later. Voice engine gives the prices.

   

5. Prepare the speech recognition model (with “Talk by voice”)

   On a Mac with 16 GB of memory or more, ASIST recommends Qwen3-ASR 1.7B, and Qwen3-ASR 0.6B with less. On Windows, it recommends 1.7B with 6 GB of memory or more on the discrete GPU, and 0.6B with less. Click “Prepare the model” to download the model (about 2.2 GB for 1.7B, about 0.8 GB for 0.6B) and run it on your computer’s GPU. Under “Details” you can choose another model; the list shows only the models that recognize the conversation language (Voice engine). You can also choose Whisper running inside the browser. On Windows without a discrete GPU, the reason appears in place of the model, so choose Whisper running inside the browser.

   On Windows, the first start of a new version of ASIST, and the first start after a GPU driver update, is when the driver builds the programs that run on the GPU. Only that time, the preparation takes longer than usual.

   

6. Choose a voice for reading aloud (with “Talk by voice” or “Type, and hear the replies”)

   Only the options that can speak the conversation language appear. For Japanese, these are Irodori-TTS, Qwen3-TTS, “macOS voice”, VOICEVOX and AivisSpeech on a Mac, and Irodori-TTS, Qwen3-TTS, “Windows voice”, VOICEVOX and AivisSpeech on Windows. Irodori-TTS and Qwen3-TTS appear on a Mac with 16 GB of memory or more, and on Windows with 8 GB of memory or more on the discrete GPU. Irodori-TTS is chosen to begin with for Japanese and Qwen3-TTS for the other languages, except Hindi and Indonesian, which Qwen3-TTS cannot speak and which start on the system’s voice. On a computer where the two do not appear, Japanese starts on VOICEVOX and the other languages on the system’s voice.

   Irodori-TTS reads Japanese only. It downloads its model (about 1.9 GB) and runs it on your computer’s GPU. It has three voices, which you can change later under “Voice” in “Speech” on the “Voice” page of the settings. The ▶ on each voice’s card plays a bundled sample, so you can compare them before the model is downloaded. The nine voices of Qwen3-TTS are chosen in the same place, with samples read in the conversation language. Qwen3-TTS downloads the 0.6B model (about 1.2 GB). On a Mac with 24 GB of memory or more, and on Windows with 10 GB of GPU memory or more, you can change it to 1.7B (about 2.3 GB) later under “Model” in “Speech” on the “Voice” page of the settings. After the change, prepare the new model with “Prepare” below it. If you choose the system’s voice, ASIST checks whether a voice for that language is installed, and if not, shows you where in System Settings or Windows Settings to add one.

   

7. Allow the microphone (with “Talk by voice” or “Talk in real time with a voice model”)

   Click “Check the microphone”, and ASIST checks that it actually receives sound. On a Mac, macOS first asks for permission to use the microphone; click “Allow”. Windows does not ask.

   

   Tip

   If you did not allow it, you can open System Settings or Windows Settings from this screen, or switch to typing. To allow it later, see Microphone permission.

8. Wait for the other models

   ASIST prepares the model for searching the memory by meaning. When you choose “Talk by voice” for a conversation in Japanese, it also prepares a model that chooses which backchannel to say, and MaAI, which detects when you have finished speaking. MaAI takes a lot of processing, so it is prepared but left off; to use it, turn on “Turn-taking (MaAI)” under “Voice” in the settings. For anything that could not be prepared, you can try again or leave it for later. Anything you leave for later shows on “Overview” in the settings, where you prepare it.

   

9. Review and start

   Look over your choices and click “Start with these settings”.

   

Once ASIST starts, click MIC OFF and talk, or type in the input box and send. How to talk to ASIST is covered in Using ASIST.

If you want to use Agent jobs, go on to Setting up the Agent.