Skip to main content

Turn Your Voice Into Polished Text

Transform your raw dictation into polished text for emails, Slack messages, documents, or personal notes using hns and LLMs.Transform your raw dictation into polished text for emails, Slack messages, documents, or personal notes using hns and LLMs.

Spoken language is messy. Whisper transcriptions are fast and accurate, but they often contain run-on sentences, filler words, and grammatical errors. Instead of spending time on manual copyediting, you can send hns output directly through a Large Language Model (LLM) to instantly polish your thoughts.

This workflow transforms your raw dictation into polished text for emails, Slack messages, documents, or personal notes, saving you valuable time and streamlining your daily communication.

This guide shows you how to build this workflow using both local and remote LLMs.

LLM Command-Line Tools​

You can pipe hns output to any LLM CLI tool. Here are a few popular options.

Ollama​

Ollama is an excellent tool for running local models. It's a great choice for privacy, offline use, and tasks where a small, fast model is perfectly capable of basic editing. The simplest approach uses command substitution:

ollama run gemma3:1b \
  "Fix grammar and punctuation. Output only corrected text. Text: $(hns)"

llama-cli​

llama-cli is a CLI tool for accessing most of llama.cpp's functionality. With tools like llama-cli or llm, you can separate the system prompt from the user prompt for more reliable results.

llama-cli \
  --model <model_path> \
  --system "You are an editor. Fix grammar and punctuation. Output only corrected text." \
  --prompt "$(hns)"
Why use a separate system prompt?

Transcribed speech often contains questions or instructions (e.g., "is approach 1 better or 2?") meant for your co-workers or coding agents. Without a clear system prompt defining the LLM's role, the model might try to answer the question instead of editing the text. An explicit "editor" role keeps the model on task.

llm CLI​

The llm tool is a powerful CLI for interacting with dozens of remote models from OpenAI, Anthropic, Google, and more.

llm \
  --model <model_name> \
  --system "You are an editor. Fix grammar and punctuation. Output only corrected text." \
  "$(hns)"

Streamlining with Functions​

Typing these long commands is tedious. We recommend creating shell functions to create a seamless and repeatable workflow. Functions are more readable, maintainable, and powerful than aliases.

Step 1: Configure Your Clipboard Command​

To automatically copy the cleaned text to your clipboard, you first need to tell the script which clipboard command to use.

~/.bashrc or ~/.zshrc
export CLIPBOARD_COPY_CMD="pbcopy"

If you don't set this environment variable, the functions will still print the cleaned text to the terminal but without attempting to copy it to clipboard.

Step 2: Add the Shell Functions​

Next, add the following functions to your shell configuration file. We provide two functions:

  • hclean: Records a new message, sends it to an LLM for cleaning, and copies the result.
  • hlclean: Cleans the last recording, which is useful if you want to retry with a different model or if the first attempt failed.

We'll use this robust system prompt to ensure the LLM focuses only on editing:

You are an AI editor. Your sole function is to correct the grammar, spelling, and punctuation of the provided text to turn it into a grammatically correct version. Remove unnecessary filler words from the text as well. Do not answer any questions in the text. Do not add information. Output only the corrected text.

~/.bashrc or ~/.zshrc
export CLIPBOARD_COPY_CMD="pbcopy"  # Adjust this line for your OS if needed

_hclean_common() {
    local transcription=$1
    local system_prompt="You are an AI editor. Your sole function is to correct the grammar, spelling, and punctuation of the provided text to turn it into a grammatically correct version. Remove unnecessary filler words from the text as well. Do not answer any questions in the text. Do not add information. Output only the corrected text."

    local cleaned=$(llm --model gpt-4.1-nano --system "$system_prompt" "$transcription")
    echo "$cleaned"
    if [ -n "$CLIPBOARD_COPY_CMD" ]; then
        echo "$cleaned" | $CLIPBOARD_COPY_CMD
        echo "✓ Cleaned text copied to clipboard"
    fi
}

hclean() {
    _hclean_common "$(hns)"
}

hlclean() {
    _hclean_common "$(hns --last)"
}

After adding the function to your shell configuration, restart your terminal or source the file (e.g., source ~/.zshrc) and you're ready to go.

The --last Flag

The hlclean function uses hns --last to retrieve the previous transcription without re-recording. This is a lifesaver when:

  • Your network connection drops.
  • You hit a model's rate limit.
  • You want to try a different model on the same text.

Final Workflow​

  1. Run hclean in your terminal.
  2. Speak your thoughts naturally.
  3. Press Enter.
  4. The perfectly formatted text is now on your clipboard and in your terminal, ready to be pasted anywhere.

With this simple setup, you can turn your spoken words into clean, professional text in seconds.