Blog · Guide

Arabic speech to text for meetings: how to test a tool

By the Echo team · Published 3 November 2026 · 6 min read

In short: Choose an Arabic speech-to-text tool by testing it on 10 to 15 minutes of your own meeting. Write a correct transcript of that section by hand, run every tool on the same audio, and measure word error rate, the share of words a tool gets wrong. Count names, numbers and English terms separately, because those errors cost the most. Published benchmarks help you shortlist, but only your own recording shows how a tool handles your room, dialect and languages.

What does accuracy mean for Arabic speech to text?

Speech to text, also called automatic speech recognition, turns recorded speech into written words. For Arabic, accuracy is harder to judge than for English, because a meeting often moves between Modern Standard Arabic, a spoken dialect and English terms inside the same sentence. A tool can be strong on one of those and weak on the others.

The standard measure is word error rate, or WER. It counts three kinds of mistake: substitutions (a wrong word in place of the right one), deletions (a word that was said but is missing) and insertions (a word that appears in the transcript but was never said). The total is divided by the number of words in a correct reference transcript. A NIST slide presentation from the TREC series describes the same formula, and it is the basis of the scoring tool SCLITE. Lower is better: a WER of 10 percent means about one word in ten is wrong.

WER has limits. It treats every word the same, so a missed “not” in “we will not sign” counts as much as a missed “so” in filler speech. It also depends on spelling. Arabic dialects have no single standard spelling, so a correct word written in another way can count as an error. Read how to write Arabic dialect in notes and transcripts before you judge a transcript on spelling alone.

Why are published accuracy figures not enough?

Vendors publish accuracy figures, and they are useful for making a shortlist. But they are measured on audio the vendor chose, often broadcast news, and on the model version of that day. A 2024 paper on Arabic telephone speech, built from about 132 hours of annotated calls covering six dialects from 13 countries, found large gaps between systems. Among the models its results table lists, the word error rate ran from 48.9 percent for Chirp to 83.8 percent for Whisper Large V1. Those were 2024 versions, and models change, so treat the numbers as a snapshot. The lesson holds anyway: the same language produces very different results depending on the tool and the audio.

A 2024 test published by the speech company SESTEK reached a similar conclusion from the other side. It found that one engine can do better than others on a specific recording and worse on another. Because SESTEK compared its own system with competitors, read its results as a reason to test, not as a verdict. Open research leaderboards, such as the Open Universal Arabic ASR Leaderboard, compare open models on several multi-dialect datasets. They are a good place to start reading, and still not a substitute for your own meeting.

How do you test a tool on your own meetings?

Choose a real section. Take 10 to 15 minutes of a meeting that matters to you, with the people, the room and the mix of languages you normally have. A clean studio clip proves very little.
Ask for consent first. Tell everyone in the meeting that you are recording, and ask before you start. The rules differ by country, so check the official guidance where the meeting takes place. This is not legal advice.
Make a reference. Write a correct transcript of the section by hand, or have a colleague check a draft. Decide before you start how you will write dialect words and English terms, and apply the same rule to every tool’s output.
Run every tool on the same file. Same audio, same length, same settings. Do not let one tool hear a cleaner copy than the others.
Score it. Compare each output with the reference and count substitutions, deletions and insertions. A short section can be scored by hand in a spreadsheet. For longer recordings, use a script that computes WER.

What should you count besides the WER score?

Two tools can have a similar WER and still give you very different notes. Count these errors separately:

Names and numbers. A wrong client name or a wrong amount in the minutes matters more than a wrong filler word.
English terms inside Arabic speech. Product names, job titles and acronyms are where a tool often breaks when the language switches. The code-switching entry explains why.
Who said what. Check the speaker labels, because decisions and action items depend on who made them. Speaker diarization is the step that separates one voice from another.
Negatives and dates. “We will” and “we will not” reverse a decision, and a wrong date moves a deadline.

How should you decide after the test?

For each tool, record three things: the WER, the number of serious errors from the list above, and how long it took before the notes were fit to send. Choose the tool with the fewest serious errors that fits your time. Then repeat the test on a second meeting, with a different speaker or dialect, before you pay for a plan. One good result on one recording is not enough.

What else should you check before you buy?

Accuracy is only one question. Ask the vendor these before you commit:

Which dialects does it name? A word like “Arabic” alone tells you little about Gulf, Levantine, Egyptian or Maghrebi speech.
Does it join the call as a visible participant, or does it record from your own device? The bot or no bot guide explains the difference.
Where is the recording kept, how long are copies held, and how do you delete them? Get the answer in writing.
How many minutes can you test for free, and what does the paid plan cost in your currency?

For a worked example of reading a tool’s language list before you test it, see Otter alternative for Arabic meetings: what to check first.

Where Echo fits

Echo is an AI notetaker for iPhone and Mac, made in the UAE by CNTXT. It works in Arabic and English, covers Gulf, Levantine, Egyptian, Maghrebi and Modern Standard Arabic, and handles recordings that mix Arabic with English. On Mac it records the microphone and the computer’s audio from the menu bar, so no bot joins the call. The free plan includes 30 minutes of recording a month, up to 15 minutes per recording. Test Echo on your own meeting like any other tool before you decide. Prices are on the pricing page.

Frequently asked questions

What is a good word error rate for Arabic meetings?

There is no universal pass mark. The right level depends on the stakes. A personal memo can tolerate more small errors than minutes that record a financial or legal decision. Compare tools on the same recording, and look at the serious errors separately: names, numbers, negatives and speaker labels. Published figures are useful for a shortlist, but they come from other audio and older model versions.

Can I test an Arabic transcription tool for free?

Many tools offer a free allowance. Echo’s free plan includes 30 minutes of recording a month, up to 15 minutes per recording, which is enough for a short test on a real meeting. Check the current limits on the pricing page before you start, because plans can change.

Do I need to tell people before I record a meeting?

Yes. Tell everyone in the meeting that you are recording, and ask before you start, even if the tool records only from your own device. The rules differ by country, so check the official guidance where the meeting takes place. This is general information, not legal advice.

Why does a tool miss English words in Arabic meetings?

When Arabic and English switch in the middle of a sentence, a tool may choose the wrong language for a word, or write an English term in Arabic letters. Test a section that contains your product names and technical terms. If a term keeps coming out wrong, fix it by hand in the notes before you share them.

Sources

Checked on 3 November 2026.

Try Echo on your next meeting. Free for 30 minutes a month.

Try Echo free