Most dictation apps stop working the moment you stop talking. You get a wall of raw text, complete with every “um,” every false start, every sentence that trailed off into nothing. Then it is on you to clean it up.

Superscribe’s smart transcription pipeline changes that. It takes raw voice and runs it through two stages: instant filler removal, then AI-powered template formatting. The result is output you can actually use without editing.

Here is how it works.

Stage 1: Filler Word Removal (Instant)

The first stage happens the moment your audio is transcribed. Before any AI touches the text, filler words get stripped out instantly.

This is not a language model making judgment calls about what to keep. It is pattern matching. Fast, predictable, and thorough.

What gets removed:

• Filler words: “um,” “uh,” “hmm,” “like,” “you know,” “basically”
• Language-specific fillers across 99+ languages: German “äh,” French “euh,” Estonian “noh,” Spanish “pues,” and dozens more
• Repeated words: “the the,” “I I,” “so so”
• Punctuation artifacts left behind after removals (orphaned commas, double spaces, trailing dots)

The key detail: this adds zero latency. It runs as a pure text transform, no network call, no model inference. By the time you see your transcription, the fillers are already gone.

For many use cases, this is enough. You dictate, the filler disappears, and you have clean text. Done.

But sometimes you need more than clean text. You need formatted output.

Stage 2: AI Template Processing

This is where the 13 output templates come in. Pick a template, and an AI model takes your cleaned transcription and reshapes it into a specific document format.

The processing is sentiment-aware. It reads the tone and intent of what you said, not just the words. A frustrated bug report stays frustrated. A casual message stays casual. The AI reformats without flattening your voice.

A spinner appears while this runs, so you know processing is happening. Depending on the template and length, it takes a few seconds.

Let’s walk through what this actually looks like with real examples.

Example 1: Raw Dictation to Professional Email

What you said:

“Hey so I wanted to follow up with the client about the logo designs, the second round is done and I think option B is the strongest but we should let them decide, also need to mention the timeline is shifted by about a week because of the feedback delay on their end”

After filler removal:

“Hey I wanted to follow up with the client about the logo designs, the second round is done and I think option B is the strongest but we should let them decide, also need to mention the timeline is shifted by about a week because of the feedback delay on their end”

After Professional Email template:

Subject: Logo Design Update - Round 2 Complete

Hi [Client],

The second round of logo designs is complete. Of the options presented, Option B stands out as the strongest direction, but we would love your input before moving forward.

A quick note on timeline: due to the delay in receiving feedback from the previous round, the schedule has shifted by approximately one week. We will send an updated project timeline separately.

Please let us know your thoughts on the designs at your earliest convenience.

Best regards

One dictation. No editing. Ready to send.

Example 2: Brain Dump to Meeting Notes

What you said:

“Okay so in the standup today Marcus said the API integration is blocked because the third-party docs are wrong, Sarah is finishing the onboarding flow should be done by Thursday, and we agreed to push the launch from March 1st to March 8th because of the API thing, oh and we need to book the demo room for Friday’s client presentation”

After Meeting Notes template:

Meeting Notes

Attendees: Marcus, Sarah

Updates:

• API integration blocked due to incorrect third-party documentation (Marcus)
• Onboarding flow on track, expected completion by Thursday (Sarah)

Decisions:

• Launch pushed from March 1st to March 8th due to API integration blocker

Action Items:

• Book demo room for Friday client presentation

The template identified attendees, separated updates from decisions, and extracted action items. All from a single stream-of-consciousness recap spoken in about 20 seconds.

Example 3: Coding - AI Prompt from Voice

This is where things get interesting for developers. The AI Coding Prompt template takes a verbal description of what you want to build and structures it into a prompt you can paste directly into an AI coding assistant.

What you said:

“I need a React component that shows a dropdown menu with search filtering, it should accept an array of options as props, highlight the matching text as you type, and call an onChange handler when you select something, also keyboard navigation with arrow keys and enter to select”

After AI Coding Prompt template:

Create a React component with the following requirements:

Component: SearchableDropdown

Props:

• options: Array of selectable items
• onChange: Callback function triggered on selection

Features:

• Dropdown menu with search/filter input
• Highlight matching text in options as user types
• Keyboard navigation (arrow keys to navigate, Enter to select)

Behavior:

• Filter options based on search input
• Call onChange handler with selected value
• Support full keyboard accessibility

The Elite Coding Prompt template goes further, outputting a JSON-structured specification with fields for context, requirements, constraints, and expected behavior. Useful for feeding into AI agents that work better with structured input.

Example 4: Stream-of-Consciousness to To-Do List

What you said:

“I need to renew the SSL certificate before it expires on the 15th, also the blog post about templates needs to go out this week, remind me to cancel the old analytics subscription we’re not using anymore, and I should probably update the dependencies in the main repo before they get too stale”

After To-Do List template:

• Renew SSL certificate (deadline: 15th)
• Publish templates blog post (this week)
• Cancel unused analytics subscription
• Update dependencies in main repo

Four items, extracted and formatted, from a single sentence you spoke while making coffee.

The Two-Stage Advantage

Why two stages instead of one? Why not just send everything to the AI?

Speed and reliability.

Filler removal is instant. If you just need clean text without formatting, you get it immediately. No waiting, no chance of anything “helpfully” restructuring something you wanted left alone.

The template stage is opt-in. Pick a template when you need structured output. Skip it when you do not. This means the default experience is fast, and the enhanced experience is there when you reach for it.

It also means the AI gets cleaner input. Removing filler before template processing means the model spends its tokens on structure and meaning, not on figuring out that “um, like, basically” should be ignored.

All 13 Templates

The full template lineup covers five categories:

Core: Super (grammar cleanup), Message (casual chat formatting), Summary

Email: Professional, Casual

Organization: Note, Meeting Notes, To-Do List

Content: Tweet/Social (using a Hook-Retain-Reward framework), Blog Post

Coding: AI Coding Prompt, Elite Coding Prompt (JSON structured), Bug Report

Each template is accessible from the settings panel, which now uses an inline expand design on both macOS and Windows.

Try It

Smart transcription is available now in Superscribe v0.2.29+. Dictate something messy. Pick a template. See what comes out the other side.

Get Superscribe at superscribe.io

Speak. Track. Bill.

Related reading

• 13 Templates That Turn Rambling Into Results
• How Superscribe Works in 3 Seconds
• Why Real-Time Dictation Feels More Reliable