Two engines, one job

Ray uses two speech recognition engines because subtitle workflows do not all need the same tradeoff.

Sunray (formerly Ray Max) is tuned for accuracy. It is the engine for low-resource languages, archival material, professional subtitle work, and any moment where the text has to be as close to the spoken audio as possible.

MoonRay (formerly Ray Turbo) is tuned for speed. It is the engine for live viewing, fast previews, and large batches where getting usable subtitles quickly matters.

The benchmark behind this post compares both engines across 98 languages using the same evaluation process for every model. The goal is not a vague claim that Ray is multilingual. The goal is a repeatable way to see how the engines behave across many scripts, accents, and language families.

Note: the charts below predate the engine rebrand, so their legends still show the former names — Ray Max is now Sunray, and Ray Turbo is now MoonRay.

How to read the numbers

The main metric is Word Error Rate, usually shortened to WER. Lower is better. A WER of 5 percent means the transcript is very close to the reference text. A high WER means the engine made more substitutions, insertions, or deletions.

For easier comparison, the same result can also be shown as accuracy:

accuracy = max(0, 100 - WER)

That second view is easier to scan, but WER remains the more useful engineering metric because it shows exactly where a model is struggling.

Overall performance

Full results: all 98 languages

High-performing languages

Most challenging languages

What we measure

• Recognition quality across 98 languages
• The difference between Sunray and MoonRay
• Languages where accuracy is already strong enough for everyday subtitles
• Languages that need more training data, normalization, or language-specific handling
• Failure cases where a model inserts more text than the source audio contains

Why this matters for subtitles

A subtitle engine is not judged only by how well it handles English. Real viewing is messier: multilingual films, code-switching, regional names, noisy recordings, and content that moves between languages mid-scene.

Ray is built for that reality. Benchmarks help us choose defaults, set expectations, and decide where the next training and evaluation work should go.

The important product result is simple: Ray can choose between speed and accuracy based on the job, and both engines keep improving as more languages move through the benchmark pipeline.