Lab · Test · 10 engines · 16 Nov 2026
Who hears Masrymixed with English?
People in Cairo and Riyadh switch to English mid-sentence. We recorded 13 real lines and ran them through 10 speech-to-text engines to see which one keeps up.
From the lab
89%
of the words, heard right by the best engine. The weakest got 32%.
The question
A voice agent is only
as good as what it hears.
A typical line: “ابعت لأسامة إن الـ Maatix app محتاج build جديد”. Arabic grammar, English nouns, a name in the middle. Most engines were built for one language at a time. Before we build a voice agent for a client, we test the engines on speech like this.
- 13lines: 10 in a quiet room, 3 with a video playing
- 10engines, each run twice
- 89%of the words heard right by the best
- 32%by the weakest
What we found · best setup of each engine
Words heard
right, of 100.
Also: how many English words and people's names came out right, and how long after the last word the text arrived.
-
ElevenLabs Scribe v3Whole clip, with a words list 89% 92% 95% 0.71 s
-
ElevenLabs Scribe v3 LiveLive, with a words list 85% 76% 73% 0.19 s
-
Google Gemini 3.5 TranscribeWhole clip
66%
43%
32%
2.59 s
-
OpenAI GPT-4o mini TranscribeWhole clip, names in the prompt
66%
52%
43%
0.93 s
-
OpenAI GPT TranscribeWhole clip, names in the prompt
65%
45%
52%
0.79 s
-
ElevenLabs Scribe v2 RealtimeLive, with a words list 64% 57% 45% 0.17 s
-
xAI Grok STT 2.0Whole clip 46% 17% 16% 0.62 s
-
Deepgram Nova-3Arabic model 38% 6% 14% 1.03 s
-
AssemblyAI Universal-3.6 ProWhole clip, with a words list 33% 18% 18% 0.52 s
-
Whisper large-v3 (on Groq)Names in the prompt
32%
24%
32%
0.53 s
What surprised us
English inside Arabic
is where most break.
- 6%English in Arabic lettersDeepgram wrote most English words in Arabic script: 6 in 100 came out right. It handled the Arabic well.
- 59 → 95A list of names helpsGiven a short list of the names we use, the best engine got people's names right 95% of the time, up from 59%.
- 85 → 84Noise filters didn't helpA video playing beside the speaker barely hurt the best engine. Its background filter changed almost nothing.
- 0.19 sLive costs a littleStreaming live gave 85% with text 0.19 s after the last word. Sending the whole clip gave 89%, half a second later.
What it means for you
Test on your
people's speech.
The engine that wins in English demos can fail your customers. A day of recording real lines and scoring them picks the right engine before you build on it. It is part of the first week of every voice project we take on.
What was recorded?
Thirteen everyday lines from one speaker, in Egyptian Arabic with English words and people's names, recorded in a browser. Ten in a quiet room, three with a video playing beside him.
How is "words right" counted?
One minus the word error rate against what was really said. Fillers and Arabic diacritics don't count. Names count in Arabic or Latin letters; brands only in Latin letters, as written.
Which setups did you try?
Every engine with and without a list of names where it takes one, live and whole-clip where it offers both, and noise filters. The table shows each engine's best setup. Every setup ran twice.
Is this a benchmark?
No. It is one voice and thirteen lines: a quick, honest test of the kind we run before a project. Your people's speech may give a different order, which is why we test it.
From the lab
30/30
Sentry and Playwright: installed in every answer.
Start here
Tell the lab
where you are.
An idea, a launch, search or AI: one call tells you the next step.
We reply within one working day.





