> Agent-readable docs index: /llms.txt. Download /docs.zip to grep all markdown files locally.

---
title: Voice Messages
description: Record a voice message in Discord and Kimaki transcribes it using your project's file tree for accuracy.
icon: lucide:mic
---

Record a **voice message** in Discord and Kimaki transcribes it, then processes it as a normal prompt. Speaking is often faster than typing a long instruction, especially from your phone.

## How it works

```diagram
  voice ──▶ transcription ──▶ "📝 Transcribed" ──▶ session
              ▲
              │ project file tree for accuracy
```

1. You record a voice note in the thread.
2. Kimaki sends the audio to a transcription model.
3. The transcribed text appears in the thread prefixed with **📝 Transcribed message:** and is sent to the agent.

## Accuracy from your project file tree

Transcription of code is hard: file names and function names are not normal words. Kimaki improves accuracy by feeding the **project's file tree** into the transcription prompt, so when you say a file path or a function name, the model recognizes it from your actual codebase.

<Aside>
  <Tip>
    Say file and function names naturally. Because the transcriber sees your file list, "open thread session runtime" is far more likely to come back as `thread-session-runtime.ts`.
  </Tip>
</Aside>

## Models and setup

Kimaki auto-detects the provider from the API key you configure for audio:

* **OpenAI** key (starts with `sk-`) → uses `gpt-audio`.
* **Gemini** key → uses `gemini-2.5-flash`.

You'll be prompted for an audio API key during setup, or you can add one later from Discord when you first send a voice message.

## Sending audio back

Kimaki can also speak: when you ask for audio of some text, the agent can generate speech with `kimaki tts` and post it to the thread. See the [Commands reference](/docs/commands) for the `tts` and `upload-to-discord` CLI subcommands.


---

*Powered by [holocron.so](https://holocron.so)*