Word Error Rate is Broken: Why Dictation Apps are Testing the Wrong Metric

The standard speech-to-text metric is Word Error Rate, or WER.

But users don’t feel percentages.

A transcript can be 97% accurate and still feel like garbage if the missing 3% is a client’s name, a critical date, a URL, or broken formatting.

For voice input, the real metric is much more binary:

• Did I have to touch my keyboard after speaking?

We’re aiming for a zero-edit commit rate, not just raw transcription accuracy.

That shift changed how we built Juno. We stopped treating the live transcript and the final text as the same thing.

• The Live HUD is there for real-time confidence while you talk.

• The Final Pass is where Juno actually commits the clean text.

From there, the local writing engine takes over based on context, stripping disfluencies, formatting messy notes, drafting a structured email, or outputting clean bullets.

The product loop is not: speech → transcript

It is: messy speech → visual preview → final transcript → writing transform → safe insertion

So if you use dictation or voice-writing tools, what is the actual make-or-break metric for you?

• Zero-edit rate: It just works, no cleanup required.

• Final insertion latency: Does it paste instantly, or are you waiting around?

• Live Transcript: Can you see what you've been speaking?

• Ubiquity: Does it work natively in every app?

• Context awareness: Does it catch your specific jargon and contact names?

• Voice correction: Can you fix a mistake without grabbing the mouse?

• Reliability: Does it ever dump text into the wrong window?

• Privacy: Is your voice processed locally, or is it hitting a cloud?