Giving TARS a Voice and a Mind
This is the moment TARS comes alive. You connect its mind, a language model, and its voice and ears, text-to-speech and speech recognition, and then you talk to your robot and it answers back. Every single piece of this is the AI assistant you built in Month 3, now living inside TARS: the same choice of a local or cloud language model, the same speech-to-text and text-to-speech, wired together by the TARS-AI software.
The mind: a language model
TARS thinks with a large language model, exactly the kind you ran on the Pi in Month 3. You choose between a local model, such as one served by Ollama, which keeps everything private and works with no internet, or a cloud model reached with an API key, which is more capable but sends your words out to a company's servers. This is the same trade-off you weighed in Month 3, privacy and offline operation versus raw power, and TARS-AI supports both so the decision is yours. Whichever you pick, the model takes what you say and produces TARS's reply in its character.
Ears and a voice
For a conversation to happen, TARS has to turn your speech into text and its text back into speech. Speech-to-text, using a tool like Whisper or Vosk, transcribes what you say so the model can read it, and text-to-speech, using a tool like Piper, turns the model's reply into spoken audio through the speaker. These are the same speech pieces from your Month 3 assistant, now serving as the ears and mouth of a robot, with a wake word or a button deciding when TARS starts listening. Speech-to-text is its hearing, the language model is its thinking, and text-to-speech is its speaking, the full loop of a conversation.
Wiring it together through configuration
You connect all of this the same way you configured the Month 3 and Month 5 projects: through the TARS-AI settings and its keys file. You choose which language model, which speech-to-text, and which text-to-speech to use, and you paste in any API keys the cloud options need, keeping those keys private and out of anything you share, just as you did before. Configuring through one place rather than editing code keeps TARS easy to adjust, and favoring a local model with on-device speech keeps the whole conversation private on your own hardware, which suits a robot that lives in your home.
Which settings to change, where the keys go, and which models and speech engines TARS-AI supports are documented and maintained by the TARS-AI Community. Follow their configuration guide, since the exact options evolve, and their Discord is a great place to compare notes on which local model or voice works best:
Configuration docs: docs-tars-ai.vercel.app · GitHub: github.com/TARS-AI-Community/TARS-AI · Discord: discord.gg/AmE2Gv9EUt
When the model and speech are configured, you launch TARS and speak to it. It hears you through the microphone, transcribes your words, the language model composes a reply, and it speaks the answer while its screen shows its state. The first time your robot answers a question in its own voice is the payoff of the entire six months, and, exactly as with the Month 3 assistant, expect to spend a little time tuning the timing and the voice until it feels right.