The material decides the number
Read speech from a benchmark corpus is far easier than a four-person meeting on a speakerphone. An engine tuned for one can look excellent on paper and unusable in practice.
Why published Hebrew accuracy figures contradict each other
You can find claims of a 1.2% word error rate and a 26% word error rate for the same engine. Both can be honest. They are measuring different things.
WER treats every word equally
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 inflates the error count
Hebrew packs into single tokens what English spreads over several words. One segmentation mistake can register as several errors, so Hebrew WER is not comparable to English WER.
Code-switching is usually excluded
Most public Hebrew test sets are monolingual. Real Israeli speech is not, so benchmark scores systematically overstate performance on actual meetings and calls.
How KolWrite evaluates Hebrew
Measured on the material customers actually bring, reported by condition rather than as one number.
By condition, not in aggregate
Clean single speaker, multi-speaker meeting, phone-quality call and noisy field recording are tracked separately, because they behave differently.
Mixed Hebrew-English included
Code-switched recordings are part of the evaluation set rather than excluded from it.
Errors weighted by consequence
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.
Model routing per recording
Audio type is detected and the transcription approach is chosen to match, rather than sending every recording through one fixed model.
Questions
What is a good word error rate for Hebrew?
There is no single figure that is meaningful across conditions. On clean, single-speaker Hebrew, strong systems cluster closely together and the differences between them rarely matter. On multi-speaker, noisy or code-switched Hebrew — which is most real-world material — the spread between systems is far wider, and that is where the choice of engine actually shows.
Why is Hebrew word error rate not comparable to English?
Hebrew fuses prepositions, articles and possessives onto word stems, so a single segmentation error can be counted as several word errors. The same underlying quality therefore scores worse in Hebrew than in English, which makes cross-language WER comparisons misleading.
How should I compare Hebrew transcription tools?
Test them on your own audio, not on a benchmark. Use a recording that is representative of your hardest case — the noisiest room, the most speakers, the most English mixed in — and compare the errors that would actually cost 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 benchmark.
Take your hardest Hebrew recording — the noisy one, with four speakers and English mixed in — and compare the output yourself.
Test it on your audio