The audio dictates the score
Someone reading aloud from a benchmark corpus is much easier to handle than four people talking on a speakerphone. Tune an engine for one, and it can look great on paper while being unusable in practice.
Accuracy
What “most accurate Hebrew transcription” should mean.
People quote headline accuracy numbers for Hebrew all the time, yet almost never compare them.
Live demo
Runs in your browser.
We explain what goes into that score, why published figures clash so much, and how KolWrite measures its own.
Why published Hebrew accuracy numbers don't match up
One report might show a 1.2% word error rate for an engine while another puts it at 26%. Both numbers can be honest because they are testing completely different things.
WER counts every word the same
Word error rate counts a missed preposition the same as a missed name or figure. For Hebrew, where a fused prefix changes meaning, that flattens exactly the errors that matter most.
Morphology drives up the error count
A single Hebrew word often holds what takes several words to say in English. When the model splits a word wrong, that one mistake can count as multiple errors, so you can't compare Hebrew and English WER scores directly.
Code-switching is usually left out
Most public Hebrew test sets only include one language. Because real Israeli speech does not stick to one language, benchmark scores consistently overstate how well a model performs on actual calls and meetings.
How we test Hebrew at KolWrite
We test on the kinds of recordings our customers actually upload and report results for each audio type instead of lumping them into a single score.
Tracked by recording type
We track clean single-speaker audio, multi-speaker meetings, phone-quality calls, and noisy field recordings separately because each one behaves differently.
Hebrew-English mix included
We keep recordings with code-switching in our test sets instead of filtering them out.
Errors weighed by impact
Names, numbers, dates and terminology are scored separately from function words. A transcript that gets every name right is more useful than one with a marginally better raw WER.
Matching the model to the audio
Instead of running every file through the same fixed model, we detect the audio type first and pick the transcription setup that fits it.
Questions
What is a good word error rate for Hebrew?
No single number holds up across every situation. With clean, single-speaker Hebrew, strong systems score very close together and the small differences between them rarely matter. The gap between systems gets much wider on multi-speaker, noisy, or code-switched Hebrew, which makes up most real-world audio. That is where your choice of engine actually shows.
Why is Hebrew word error rate not comparable to English?
In Hebrew, prepositions, articles, and possessives attach directly to the word stem. Because of this, one splitting mistake can show up as several word errors. A transcript of the exact same quality will get a worse WER score in Hebrew than in English, so comparing scores across the two languages is misleading.
How should I compare Hebrew transcription tools?
Skip the benchmarks and test them on your own audio. Pick a recording that represents your toughest case, with the noisiest room, the most speakers, and the most English mixed in. Then compare the mistakes that would actually take you time to fix.
Does accuracy differ between Hebrew audio types?
Substantially. Recording conditions, number of speakers, microphone distance and how much English is mixed in all move accuracy more than the choice between leading engines does on clean audio.
Run your own test.
Grab your toughest Hebrew recording, the noisy one with four speakers and English mixed in, and compare the output yourself.