Willow Zero-Edit Benchmark 2026: What the Score Means

Imagine this: a dictated supplier email looks finished, with clean sentences and the right tone. Then you notice that "we can't confirm delivery" became "we can confirm delivery." The punctuation is fine, but the message still can't be sent. That is the question behind the Willow zero-edit benchmark: What counts as finished dictation?

A score means more when you know which changes reviewers considered necessary. Willow's October 4, 2026 benchmark reports a 95% English zero-edit rate, compared with 90% for Wispr Flow. The report covers internal testing from September, conducted in a quiet room with a close-talk USB headset.

Two reviewers scored the outputs. These are vendor-reported results, not an independent ranking. We make DictaFlow and did not run this test.

The useful buying question isn't whether a chart looks impressive. It's whether "no edits" means the same thing to the evaluator and the person sending the draft.

Decide what counts as an edit before comparing scores

For a small business, a dictated message can create a promise. Consider this invented test sentence: "The replacement is available, but do not authorize the shipment until the revised quote is approved." An evaluator should check the condition, not just whether the paragraph reads smoothly. Removing "until" would change the instruction.

Changing "approved" to "reviewed" would change it too. A clean rewrite may still need correction. Now consider a harmless preference.

You like "Thanks" at the end of an email, but the tool produces "Thank you." Would you edit that every time? If you would, your personal no-edit score should reflect it. If you wouldn't, don't mark it as a failure simply because it differs from a reference draft.

Write these rules down before opening a second app. Otherwise, you may excuse one product's wording and penalize another's for making the same change. Brand names shouldn't determine how strict the review is.

Apple's research on transcript readability draws a useful distinction. Its HEWER metric separates major errors from minor differences and focuses on meaning, readability, and proper nouns. The research covers podcast transcripts, not rankings of desktop dictation apps.

Still, it shows why treating every text difference as equally important can hide the errors readers care about. Use a similar distinction in your own notes, but don't claim to have reproduced Apple's metric. Mark a changed business instruction as a required correction.

Put optional style preferences in a separate column. You need to track both, but they answer different questions.

Keep the unit of work fixed

A no-edit score counts finished outputs, not individual words. That makes the boundaries of each output important. A one-sentence reply and a long proposal paragraph are different units of work.

Take this invented customer update: "The inspection is complete. The warranty decision is still pending. Please keep the original part until we send return instructions." You could dictate it in one recording or split it into separate recordings.

If one sentence needs fixing, the first approach gives you one failed message. The second gives you a mix of clean and failed outputs. The wording may be identical, but the count now reflects a different way of working.

For your comparison, choose the unit you actually use. If you usually dictate a full reply, score full replies. If you build messages one sentence at a time, score those bursts.

Don't quietly change the unit for the product that looks better with shorter recordings. Record the draft length and review policy alongside the score. The Willow zero-edit benchmark methodology describes passages of 150 to 200 words.

Its number shouldn't be treated as a measured result for your brief customer replies.

Word error rate asks a different question. Hugging Face's ASR evaluation guide explains that WER counts substitutions, deletions, and insertions against a reference transcript. It helps test recognition accuracy, but it does not replace deciding whether a complete business message needs editing.

Record why you changed the text

For a practical trial, use fictional business material instead of private customer records. Prepare a few drafts that you can judge consistently. Include one instruction with a condition and one message that opens with uncertainty.

For example: "We may be able to fit the repair in next week. Please wait for our confirmation before bringing the equipment." Treat "may" and "wait for our confirmation" as protected meaning. The text shouldn't become a confirmed booking just because a cleanup model favors shorter sentences.

Add a note to each result explaining the required edit. "Changed a tentative booking into a promise" tells you more than "bad accuracy." "I prefer a shorter greeting" belongs under preference, not factual damage. If possible, review a saved copy before checking which app produced it. You won't have a controlled laboratory study, but you can reduce the urge to defend a favorite product.

Save the original output too. Once you fix the draft, you lose the evidence of what needed work. Also record the review time.

A message may need no correction and still require careful reading before approval. Zero edits should never mean zero review, especially when a draft includes a deadline, permission, or commitment.

Don't invent one passing score for everything. A wrong greeting and a reversed shipping instruction don't have the same practical impact. Keep serious mistakes visible instead of hiding them in one average.

Apply the same standard to DictaFlow

DictaFlow is one candidate for this trial, not a winner established by Willow's chart. The published comparison doesn't include a DictaFlow result, so we shouldn't fill that gap with a claimed score. On desktop, hold your shortcut, speak, and release it to insert editable text at the cursor.

Custom vocabulary can help with recurring names and specialist terms. Neither feature removes the need to check whether the draft preserved your instruction. DictaFlow's processing options include Local Offline on supported devices, along with cloud features.

Choose the mode before testing, then label the result. Testing one mode doesn't establish the quality of every model or setting in the app. Start with the getting-started guide and use the free tier for a small, repeatable sample.

The comparison page covers broader workflow choices. Keep those choices separate from your edit rubric. Platform availability or remote-desktop compatibility can't tell you whether a sentence retained its meaning.

The Willow zero-edit benchmark gives you a good reason to ask a more precise buying question. Before trusting a percentage, decide which edits your work actually needs. Then inspect the unedited drafts. The paragraph that looks best may still be one you can't send.