On-Device Speech Recognition Benchmark on a Mac

Calculated from the rows

Every statistic is calculated by code from the rows above. Open one to see the exact values it summarises.

• 4.74 percentIn the on-device speech recognition word error rates on FLEURS dataset, the 9 published figures for “Parakeet TDT v3” have a median of 4.74 percent, ranging from 2.52 to 19.83.Calculated from other facts
• 5.08 percentIn the on-device speech recognition word error rates on FLEURS dataset, the 11 published figures for “Whisper large-v3-turbo” have a median of 5.08 percent, ranging from 2.53 to 23.51.Calculated from other facts
• 6.66 percentIn the on-device speech recognition word error rates on FLEURS dataset, the 6 published figures for “Apple SpeechTranscriber” have a median of 6.66 percent, ranging from 4.15 to 100.Calculated from other facts
• 26 data pointsThe on-device speech recognition word error rates on FLEURS dataset holds 26 data points, each a published figure traced to its source.Calculated from other facts

How this dataset is built

What counts as a row

Our own measurement: each engine transcribes the same 100 recordings per language from the FLEURS test set (Google, 2022), chosen by a fixed rule before any engine ran, on one Mac with nothing sent over the network. Every engine gets the language it is asked to recognize.

What does not

Cloud services (they cannot run on the device) and languages an engine does not support.

How rows are found and checked

The recordings are the FLEURS test split at a pinned revision, 100 per language at evenly spaced positions. Each engine's text and FLEURS's written sentence are normalized the same way (Unicode NFKC, lowercase, words as runs of letters and digits) and compared word by word; the word error rate is the total of substitutions, deletions and insertions over the total of reference words. Every transcript, every timing and the SHA-256 of every recording are published in the raw record, so anyone can recompute every number.

What the data cannot tell you

Read sentences from Wikipedia in quiet conditions are easier than dictation, conversation or noise. Numbers and spelling variants are not normalized, which counts '25' against 'twenty-five' as an error for every engine alike. One Mac, one run; we make Yap, which uses Parakeet, so we publish everything needed to check us.

The whole method, from search to calculation, is on the research data page.

Cite and reuse

You may copy, share and adapt this dataset for any purpose, including commercially, if you credit it with a link. Quoted passages remain their authors' words: cite the original source for those.

Yap (2026). On-device speech recognition word error rates on FLEURS [Data set]. https://yapapp.cc/data/on-device-asr-benchmark

License: CC BY 4.0.

Pages built on this data

• Speech Recognition Accuracy: Word Error Rates

More datasets

• Typing Speed Statistics: Average WPM by Group14 data points
• Speech vs Typing Speed: Text Entry Rates Compared57 data points
• Speaking Rate Statistics: Words per Minute7 data points
• Speech Recognition Accuracy: Word Error Rates597 data points
• Speech Recognition Speed: Real-Time Factor Data24 data points
• How Many Languages Speech Recognition Supports45 data points
• Speech-to-Text API Pricing: Cost per Hour of Audio34 data points
• Dictation Software Prices: What Each App Costs37 data points
• How Long Voice Assistants Keep Your Recordings38 data points
• Repetitive Strain Injury Statistics: How Common Is RSI18 data points
• Disability Statistics: Conditions That Make Typing Hard22 data points
• Voice Assistant and Voice Search Usage Statistics2 data points