FLEURS short dictations
243 prompted recordings across 9 languages.
View the FLEURS datasetWasper voice engine benchmark · Wasper 1.8.0 · August 2026
Accuracy and speed don’t have one winner.
Your language and workload change the answer. Explore 3,960 measured transcriptions across Cohere, Whisper, and Parakeet — then inspect the exact recordings behind every summary.
- Languages
- 9
- Original recordings
- 264
- Voice engines
- 3
Short dictation
Performance map
Lower word error rate is better. Higher processing speed is better. Choose a flag to focus that engine and language. Every exact processing speed was measured on MacBook Pro · Apple M1 Pro · 16 GB memory · macOS 14.7.3. *
- Cohere — gold flag wrapper
- Whisper — blue flag wrapper
- Parakeet — olive-green flag wrapper
Processing speed (x real time)20x real time means 1 hour of audio is processed in about 3 minutes.Word error rate (%)
Focus or hover a point for exact values.
Choose a workload
See how each engine handles the way you speak.
Quick dictations and long recordings challenge speech engines differently.Choose the workload closest to yours and every result updates.
Interactive explorer
Compare the exact workload you care about.
Choose a voice engine, then select the languages that matter to you.Every filter updates the full comparison together.
Voice engine
Language
Short dictation · All engines · English
Voice engine comparison
For English short speech on MacBook Pro · Apple M1 Pro · 16 GB memory · macOS 14.7.3, Cohere had the lowest WER; Parakeet was fastest.
| Voice engine | Transcription quality WER counts incorrect words. CER counts incorrect characters. Lower percentages mean a more accurate transcription. | Processing speed |
|---|---|---|
| CohereLowest WER | Chosen language4.8% WER2.4% CER | Chosen20.3×* Automatic language detection is not available for Cohere. |
| Whisper | Chosen language7.4% WER5.0% CER Same WER whether the language was chosen or detected automatically. | Chosen5.3×* Auto detection4.0×* |
| ParakeetFastest | Chosen language5.7% WER3.1% CER Same WER whether the language was chosen or detected automatically. | Chosen63.5×* Auto detection64.7×* |
Inspect the 27 English recordings
RecordingLengthSourceDetails
Sources and test setup
Built from trusted, credited sources.
Every benchmark recording is clearly credited. Follow it back to the original dataset or publisher.
Long recordings
21 complete talks, speeches, readings, and addresses from their original publishers and datasets.
- Test Mac
- MacBook Pro · Apple M1 Pro · 16 GB memory
- Operating system
- macOS 14.7.3
- Power
- Connected to AC
- Method
- Every engine received the same audio for 3 passes.
- Speed
- N× means N seconds of audio processed per second in this direct-server benchmark.
- Benchmark date
- August 16, 2026
Primary evidence SHA-256724014a21cb0fe82d24e720a7a4810013e9ff612d39aa68c35e8b536225b8550
Public data SHA-2566cc07e61964684da6b467ebfe055e3b705771ab72733f0fc62834314b8696767
Choose the engine that matches your work.
Wasper lets you change voice engines without changing how you dictate. Start with the benchmark, then decide what feels right on your Mac.