Mozhi

Starting the router

Speak, naturally.

Tamil. Malayalam. English.
Your language. Your level of respect.

Ready when you are

Ask for something. Mozhi decides whether to do it.

Preferences and saved items

Decision journal

    Words in.
    Possibilities out.

    Ask something. See what Mozhi heard,
    what it chose, and what happens next.

    Example decision

    play some jazz in the kitchen

    1. Heard
    2. Routed
    3. Not done

    That sounds like a request to play music, which I cannot do.

    A useful assistant knows its limits.
    Every action comes with a reason.

    What is Mozhi?

    Mozhi (மொழி, മൊഴി: "speech" or "language" in Tamil and Malayalam) is a free, open-source voice assistant for Tamil, Malayalam and English. It checks before it acts: a small trained router decides in about 2 milliseconds whether to check the weather, set a reminder, save a note, tell a story, recite a verse or read the news, and it declines requests it cannot do.

    Does it need a large language model?

    Not to choose what to do. On 2,805 test requests the router scored 0.941 action macro-F1 with 0.8% false actions, against 0.879 for the best LLM prompt on a 539-request subset, and it is roughly 500 to 1,500 times faster. A language model is only used for open questions, through your own key or a local gateway.

    How does the respect meter work?

    Five levels change how Mozhi addresses you: from a close friend (da, di) through family respect (anna, akka, chetta, chechi) to polite and deferential forms (sir, ayya, muthalali), together with the polite or informal "you".

    Which languages work today?

    English is measured and finished. Tamil and Malayalam script are routed by a first multilingual model that scores 0.82 and 0.81, below its 0.85 target, so it is still being improved. Tanglish and Manglish requests work best when mixed with English words. The Tamil and Malayalam wording is a draft awaiting native-speaker review.

    Source code and full evaluation on GitHub