Dictation accuracy benchmark: six Mac dictation apps on Indian English and Hinglish

On the same eight clips of everyday Indian English and Hinglish, Caddie Scribe made the fewest word errors (7.0%), Wispr Flow was close behind (7.9%), and Apple's built-in dictation got about one word in four wrong (25.5%).

Measured 13 October 2026 by the team that makes Caddie. 8 clips, one speaker, about 70 seconds of audio. Download the data (CSV)

Key findings

1. In Caddie's test of six Mac dictation setups on the same eight clips, Caddie's Scribe mode made the fewest word errors: 7.0%.
2. Wispr Flow had a 7.9% word error rate on the same clips, less than a point behind Caddie Scribe, which is within this test's noise.
3. Apple's built-in macOS dictation got about one word in four wrong (25.5% word error rate) on everyday Indian English and Hinglish speech.
4. Apple dictation was the quickest to show text, a median 0.23 seconds after the key was let go, against 1.01 seconds for Caddie Scribe.
5. Sending the whole clip once the speaker stopped, instead of streaming it, cut ElevenLabs Scribe's word error rate from 13.3% to about 7%.
6. Compressing the audio as AAC while the speaker talked made the upload 7 times smaller (2.3 MB to 0.33 MB) and cut Scribe's wait from 1.24 seconds to 0.91 seconds.
7. A relay server in Mumbai cut the network round trip for a speaker in India from 0.54 seconds to 0.29 seconds.

Word error rate by app

Lower is better. Word errors are the wrong, missing and extra words, as a share of the words the speaker meant to type.

The data

Wait is the median time from letting go of the dictation key to the text appearing in the app. Download this table as CSV.

Method and sample

• Sample. Eight everyday clips from one speaker, about 70 seconds of audio in all: Indian English, Hinglish (Hindi and English in one sentence), tech terms, people's names, a self-correction, an email address and a number.
• Same sound for every app. The clips were played into each app through a virtual microphone, so every app heard exactly the same audio.
• Scoring. Wrong, missing and extra words were counted against what the speaker meant to type, and divided by the number of words meant.
• Cleanup counts. Caddie's three modes were scored after their cleanup step, because that is the text a user gets. Apple dictation has no cleanup.
• Script. Kivi writes Hindi in Devanagari, so its output was converted to Roman letters before scoring.
• Setups. Six setups, each as shipped on 13 October 2026: Caddie Scribe, Default and Turbo, Wispr Flow, Kivi by Sarvam, and Apple dictation on macOS.

Limits

How to cite

You're welcome to use these numbers, under CC BY 4.0. Please link to this page.

Related