How accurate is speech recognition in French?

French is the least accurate of the five big Western European languages we tested, and the three on-device engines sit closer together here than in any other of them. In our October 2026 benchmark on a Mac with an M2 Pro, Parakeet TDT 0.6B v3 had a word error rate of 5.29 percent ¹, Apple SpeechTranscriber had 6.59 percent ², and Whisper large-v3-turbo had 6.62 percent ³. Each score is the share of wrong words on 100 French recordings from the FLEURS test set.

Apple's built-in model matches Whisper in French and is much faster

In most languages Apple's model trails the open engines. In French it does not: its error rate is within a few hundredths of a point of Whisper's, and Parakeet leads by more than a full point. Apple's speed advantage is large. It ran about 46 times faster than real time ⁴, against about 18 times for Parakeet ⁵ and about 12 times for Whisper ⁶. See the speed dataset for all languages.

Our measurement of all three engines in French

Lower is better in this table.

The transcripts behind each row are on the on-device benchmark page.

Published results confirm French is harder than Spanish or Italian

On FLEURS French, the Voxtral paper reports 4.03 percent ⁷ for Voxtral Small, 4.51 percent ⁸ for GPT-4o mini Transcribe and 5.55 percent ⁹ for Whisper large-v3. NVIDIA reports 5.15 percent ¹⁰ for Parakeet. The older Whisper large-v2 scored 8.3 percent ¹¹, so newer models have improved.

French also reacts strongly to the kind of audio. On Common Voice French, which is crowd-sourced microphone speech, Gemini 2.5 Flash had 11.86 percent ¹², nearly double its FLEURS result of 6.17 percent ¹³.

Showing 8 of 31 rows. See every row in the dataset.

All published rows are on the published accuracy page.

French ranks seventh of 25 by its best engine

Taking the best engine in each language, French comes seventh lowest of the 25 languages in the benchmark. Italian (2.52 percent ²¹), Spanish (2.53 percent ²²), English (3.92 percent ²³) and German (4.84 percent ²⁴) all do better. French is still far ahead of the languages at the bottom of the table.

A plausible cause is French spelling: many words sound the same but are written differently, so an engine can hear correctly and still write the wrong word. The benchmark does not isolate this.

What to expect when you dictate in French

Expect a few more corrections than in Spanish or Italian, mostly homophones, agreement endings and elisions. Since Apple's model is competitive here, the built-in Mac dictation is a reasonable first try, and Parakeet is the choice if you want the lowest error rate. Spoken punctuation works with French words: Apple's French help maps "deux-points" to the colon ²⁵.

Our guide to European languages in speech to text helps you choose an engine, and the languages overview covers other languages.

How to choose an engine for French

French is the one language here where speed can reasonably decide the choice. Apple's model is as accurate as Whisper in this test and several times faster, so it is a sound default for everyday dictation. Choose Parakeet when you want the lowest error rate, for example for longer documents with names and accented words. Before you commit, speak a minute of your own typical text into two engines and count the corrections: your accent, vocabulary and microphone will matter more than the small gaps in the table above. The published accuracy page shows how cloud services compare.

What these numbers do not show

The recordings are people reading prepared sentences in a quiet room, scored on one Mac. Canadian and African French, spontaneous speech, noise and specialist vocabulary are not tested, and 100 recordings is a small sample. For the measure itself, see the guide on how accurate speech recognition is. Compare with German and Italian.