Whisper vs Parakeet: a real hearing test on a laptop

My voice assistant, Edison, hears me on my own laptop. No cloud, no subscription for the ears. For the first version I used OpenAI's Whisper (the small.en model), because it's the default everyone reaches for. But it was slow, and it kept mishearing the names I say most: my own products.

So I ran a proper test: 30 real commands, recorded in my own voice, scored on two local models. Here's what happened.

The setup

The clips: 30 short recordings of things I actually say to Edison. "Hey Edison, what's on my calendar tomorrow?" "Tell the GlowBop team to fix the wall offset." "Check Kalodata for the top selling products this week." "Never mind, stop."
The models: Whisper small.en (through faster-whisper) and NVIDIA's Parakeet v3, both running on the laptop's processor, no graphics card.
The same fix-up for both: after hearing, both get the same spelling list, which corrects names a model is known to mangle.
What I scored: how long each clip took, and whether 17 brand and product names came out right.

The results

	Whisper small.en	Parakeet v3
Average time per clip	12.10 s	1.86 s
Slowest clip	28.80 s	3.22 s
Brand names right	15 of 17	17 of 17

Parakeet was about six and a half times faster on average, and nine times faster on the slowest clip. It also got every brand name right.

Where the difference showed

Two examples tell the story.

"No, I said GlowBop, not the shop." Ironically, the sentence where I correct the assistant. Whisper heard "glowce up", "global" and "glow Bob" on different tries. Parakeet heard "GlowBop" every time.

"Check Kalodata for the top selling products." Kalodata is a product research tool, and to Whisper it was "callo data" on every try. Parakeet got "Kalodata" most of the time. (This is exactly what the spelling list is for: "callo data" now becomes "Kalodata" whichever model heard it.)

Parakeet wasn't perfect. It wrote "nanolief" for a light brand a few times, where Whisper got it right. Every model has its blind spots, which is why the spelling list exists.

A note on my "word error" numbers

If you run tests like this, learn from my mistake. Most of my clips ended up with the sentence said two to six times in a row. The scorer compared each clip to one copy of the sentence, so every repeat counted as "extra words" and the word error rates came out as nonsense: hundreds of percent for both models. They're useless, so I've left them out. Speed and brand names are the honest numbers here.

Next time: one sentence per clip.

Why I haven't switched yet

The obvious move is to switch Edison's ears to Parakeet. I haven't, for one reason: the TV.

All 30 clips were recorded with the TV off. When I tried a stand-in test clip mixing my voice with TV sound, Parakeet happily wrote down what the TV said instead of what I said. In my living room the TV is on a lot, so that matters more than speed.

The next step is recording the same sentences with the TV on, and scoring again. Until then, Edison runs on a smaller Whisper model (base.en), which hears an across-the-room phrase in about 1.6 seconds: faster than small.en, a little less accurate.

What I'd tell anyone building local voice AI

Test with your own voice and your own words. Benchmarks don't know your product names.
Count brand names separately. They matter most and break first.
Add a spelling list. It's the cheapest accuracy boost there is.
Test in the real room. TV on, music on, someone talking in the kitchen.
One sentence per clip, or your error rates will lie to you.

Hear the result

You can talk to Edison in the demo on the home page. If you want your own, sign up to get Edison: you get him first and help decide what he learns next, including which ears he ends up with.