We publish accuracy tests for 36 languages, on FLEURS, the public benchmark researchers use to compare speech models, and on real-world recordings with background noise and accents. V4-Pro has the lowest error rate of the tested providers in 25 of the 36 FLEURS languages. Accuracy varies by language and recording quality.
See the test resultsThe benchmarks page shows error rates for 36 languages on FLEURS and on internal real-world recordings with background noise and accents.
Look up your language there before you upload anything. Accuracy varies by language and recording quality.
Results are listed language by language: word error rate for most languages, and character error rate for Chinese, Japanese, Khmer and Thai. You can check the numbers instead of trusting one headline figure.
Each language compares V4-Pro with Scribe V2, AssemblyAI Universal-3.5 Pro, Deepgram Nova-3 and Whisper Large V2, scored the same way on the same clips.
The methodology and FLEURS audio samples are on the same page.
| Mode | Usage |
|---|---|
| V4-Flash | Unlimited, every plan |
| V4-Pro | Previews on Free, full on Pro |
V4-Pro is our most accurate mode. Free accounts get V4-Pro previews every week, and paid plans transcribe whole recordings in V4-Pro.
V4-Flash is unlimited on every plan, so you can keep transcribing if you use up your V4-Pro allowance.
People who use Cockatoo most often mention accuracy on difficult accents, speed and ease of use.
Your recording still matters. Strong dialects, poor microphones and background noise all raise the error rate, which is why we publish real-world results next to FLEURS.
Relative to recording length
V4-Flash transcribes at up to 120x real time. Processing time depends on the recording and how busy the service is, so a long file takes minutes.
V4-Pro runs at about 20x real time.
Cockatoo detects the spoken language before it transcribes, or you can set it for each file. It can also handle recordings where people switch between languages.
Right-to-left languages such as Arabic and Hebrew display correctly in the transcript.
Cockatoo does both jobs. It turns the recording into a transcript, then translates the transcript into another language.
DeepL translates text, but it doesn't transcribe audio.
Both transcribe uploaded audio and video. Cockatoo also publishes per-language accuracy results, translates documents, and has an iOS app for recording and dictation.
TurboScribe is built around transcription volume. On Cockatoo, the free plan already gives you unlimited V4-Flash transcripts with no daily limit.
