Speech benchLab

Arabic voices, timed side by side.

Before choosing a voice, we send the same sentence to each Arabic text-to-speech service and measure what comes back.

7Arabic voice services compared66 requests timed

04Speech bench

Speech

We send the same Arabic sentence to several voice services and record how long each request takes.

An illustration of what we measure on every request: when the sound starts, and when it ends.

7services compared
66requests timed

Text to speech

The same sentence goes to each service. We time the first sound, how fast the audio is made, the slow cases, and whether speech starts before the whole file is finished.

Speech to text

Interviews and recordings become timed text a person can correct and search. In the conversation lab you speak, read what was heard, then hear the reply.

Some of the Arabic voices we run and compare

Voxtral
CosyVoice
Chatterbox
Qwen
Fish Speech

What we measure

When the first sound arrives
How fast audio is made
The slow cases, not only the average
Whether speech can start before the whole file is finished

What we keep

The audio, the history of each run, and a baseline to compare the next one with.

All benchmarks