Written with AI assistance from the source video. The video is the authority.
Log in to leave a comment
No posts yet
If you are a freelancer paying $20 every month for AI services, $240 disappears annually. The moment you upload a manuscript to the cloud, there is a risk that your business ideas will be used as training data by the service provider. By running AI directly on your own computer, your subscription fee becomes $0 and your data never leaves your machine.
First, download and run the installation file from the official Ollama website. Next, install Fluid Voice and enable microphone and accessibility permissions in your system settings. Now, check if voice-to-text conversion works even with the internet turned off. You can write on an airplane without a network connection, with a response time of around 100ms.
In macOS System Settings, click on Sound and open the Input tab. Set your microphone input gain to a fixed 70%. If the sound is too low, it won't be recognized; if it's too loud, noise will be mixed in.
In the Fluid Voice settings, set the right Option key as your global shortcut. You must enable 'hold-to-speak' mode. It recognizes voice only while the key is pressed, and text is inserted immediately upon release. There is no need to click the editor with your mouse; the text is typed directly where the cursor is placed.
Dictation inevitably includes fillers or typos. By using the 3B model llama3.2:3b, you can correct grammar and produce clean sentences in less than a second.
Write the following script to call your local server:
`python
import requests
import json
def clean_text(text):
url = "http://localhost:11434/api/generate"
payload = {
"model": "llama3.2:3b",
"prompt": f"다음 텍스트에서 문법 오류를 수정하고 필러를 제거해: {text}",
"stream": False
}
response = requests.post(url, json=payload)
return response.json()['response']
`
Executing this code eliminates the 30 minutes you spend every day on fixing typos and formatting.
If you run multiple large models on Apple Silicon, your RAM will run low and your computer will slow down. In a 16GB RAM environment, you should use the llama3.2:3b model in a 4-bit quantized state to keep the system running smoothly.
Create an ollama_env.plist file in the ~/Library/LaunchAgents/ directory and fix your environment variables. Setting OLLAMA_NUM_CTX to 4096 will allow you to process long texts without issues. Create a symbolic link for the model storage path to an external SSD instead of your system disk. This creates a solo workspace that maintains system performance in the long term while protecting your data security.
