transcriptionparakeeton-device

Three Layers of Transcription

The same voice memo, transcribed three different ways. From Apple Speech on iOS to Parakeet v3 with post-processing — each layer improves readability while preserving the speaker's intent.

The setup

A two-minute voice memo recorded on iPhone. Stream-of-consciousness, the kind of thing you capture while walking before an idea disappears. Natural speech with fillers, false starts, and mid-thought corrections.

We ran it through three transcription layers:

  1. Apple Speech — iOS on-device recognition (what you get when you record in the app)
  2. Parakeet v3 raw — NVIDIA's ASR model running locally on Mac, no post-processing
  3. Parakeet v3 + post-processing — Same model, with dictionary-based filler removal

Each layer builds on the previous one. The diffs show exactly what changes.

Layer 1 → 2: Apple Speech → Raw Parakeet v3

The model swap. Same audio file, different recognition engine. Apple's on-device speech recognition vs NVIDIA's Parakeet running locally through TalkieEngine.

What changed

Punctuation surgery. Apple wraps fillers in commas: , um, and , uh,. Parakeet leaves them bare. This matters because comma-wrapped fillers create false pauses when you read the transcript back.

Sentence flow. Apple breaks mid-thought with periods: needs. And then becomes needs, and then. The thought continues — the punctuation should too.

Capitalization. Apple capitalizes after false sentence boundaries: way, It would becomes way it would.

Compound words. fine tuned becomes fine-tuned. Small but correct.

The word count is identical (238). The content is the same. But the Parakeet transcript reads like a more faithful representation of continuous speech.

Layer 2 → 3: Raw Parakeet → Post-Processed

Same model, now with dictionary-based post-processing. This is the layer that strips filler words.

What changed

Pure subtraction. The only change is removing um, uh, and Um. Thirteen instances. No other words, punctuation, or structure touched.

Why not strip fillers in the model? Parakeet transcribes what it hears. The fillers are real — the speaker said them. Post-processing is a separate concern: what you want to read versus what was said. Keeping these as separate layers means you can always get back to the raw transcript.

The numbers. 238 words down to 225. A 5% reduction for a 100% readability improvement.

The stack

Three layers, each doing one thing well:

  1. Better model — Parakeet v3 gets the words right, punctuates naturally, handles compound words. Runs entirely on-device through TalkieEngine.

  2. Post-processing — Dictionary-based filler removal. Surgical, predictable, reversible. The raw transcript is always preserved.

  3. Local-first — All of this runs on your Mac. No cloud API, no latency, no data leaving the device.

The gap between "good enough" and "actually readable" transcription isn't one big leap. It's small, stackable improvements — each one easy to understand, easy to debug, and easy to reverse.