Hebrew conversation analytics: why most tools fail, and what it really takes

    Basics5 min readPublished

    TL;DR

    • Most major conversation-intelligence tools were trained on English. On spoken Hebrew they reach 70%–85% transcription accuracy, not enough for reliable analysis.
    • The difficulty is more than language: call-floor slang, overlapping speech, Hebrew-English switching mid-sentence, diverse accents.
    • Below 90% accuracy, call scores and compliance checks become guessing. Hebrew-first tools reach over 95%.
    • The only test that counts: run the tool on your own real calls and read the transcript against the recording.

    On paper, every international tool "supports Hebrew." In practice, a manager running real call-floor recordings finds transcripts with holes, sentences attributed to the wrong speaker, and analysis that misses the point. This gap is not a bug, it is architecture.

    Why Hebrew is hard for generic models

    • Training data: large speech models saw orders of magnitude more English. Spoken Hebrew barely exists in their data.
    • Rich morphology: one Hebrew word folds prefixes, gender and tense together. A one-letter error changes the meaning.
    • Code-switching: mixing Hebrew and English in one sentence is normal on an Israeli call floor and a nightmare for monolingual models.
    • Pace and overlap: Israeli sales calls are fast, with interruptions that strain speaker separation.

    What happens when transcription is weak

    Transcription is the base everything rests on. At 85% accuracy, one word in seven is wrong. Script-adherence checks miss clauses that were said and flag ones that were not. Half the objections are never detected. The call score turns from a metric into a guess, and the team stops trusting the system within a week. Rule of thumb: below 90% accuracy automated analysis does more harm than good, because it produces false confidence. Above 95% you can build QA, scoring and coaching on it.

    What it takes to do it right

    A tool that truly works in Hebrew is built Hebrew-first: a speech engine trained on real spoken Hebrew calls, diarization calibrated for Israeli speech pace, and an analysis model that knows local call-floor phrases. Saleso was built exactly this way and transcribes Hebrew and English at over 95% accuracy.

    How to test a tool before buying

    • Take 5 of your real recordings, including a noisy call and a heavy accent.
    • Compare transcript to recording, sentence by sentence. Count meaning-changing errors.
    • Check speaker separation: does the system know when the rep spoke and when the customer did?
    • Ask for analysis against YOUR script, not a generic demo report.

    Frequently asked questions

    What transcription accuracy should we demand in Hebrew?

    Over 95% on your own real calls. Below 90%, any analysis built on the transcript is unreliable for QA or scoring.

    Do Gong or Chorus work in Hebrew?

    The large international tools were built for English, and their spoken-Hebrew support is partial. Before deciding, run them on your own recordings and compare transcripts, sentence by sentence, against the audio.

    What about calls that mix Hebrew and English?

    Code-switching is the hardest test case. A tool trained on real Israeli calls handles it by default, because that is how an Israeli call floor talks.

    Instead of reading about it, see it on one of your own calls.