A user sent us feedback three weeks ago. They'd been converting Farsi articles and the audio was—in their words—"painful to listen to." The pronunciation was wrong. The rhythm felt mechanical. Words got stressed in bizarre places.
They were right. Our system was using the same voice engine for everything, and while it handled English decently, it butchered most other languages.
We just fixed that.
What changed
The system now detects what language you're writing in—takes about five milliseconds—and routes to whichever voice provider does the best job for that language.
English articles get one engine. Everything else gets a new engine, which supports over a hundred languages and actually understands how they're supposed to sound.
You don't touch any settings. You don't pick a preference. It happens automatically the moment you send an article.
Why this matters if you do not listen in English
Most text-to-speech was built English-first. The training data was English. The quality benchmarks were English. Other languages got added later, almost as an afterthought.
The result: if you listened to TTS in Arabic or Korean or Hindi, you could hear it. The intonation was flat. Important words got swallowed. The whole thing sounded like a robot reading phonetically without understanding what any of it meant.
That's what was happening with OutloudAI. English users got decent audio. Everyone else got a significantly worse experience, and we were charging the same price for both.
The solution wasn't asking people to pick their language manually—nobody wants another setting to configure. It was making the system smart enough to figure it out and route appropriately.
What you will hear now
If you listen in English, nothing changes. Same voice quality as before.
If you listen in Farsi, Arabic, Chinese, Japanese, Turkish, Hebrew, or any of the other languages we support—you'll notice immediately.
Better pronunciation. Natural pacing. Emphasis that makes sense. The audio sounds like someone who actually speaks the language, not a machine approximating sounds.
One early tester converted a Chinese news article before and after the update. Before: choppy, wrong tones, barely intelligible. After: smooth, correct pronunciation, actually pleasant to listen to.
That's the gap we're closing.
The languages that work better now
The new voice engine handles these particularly well:
Arabic, Bengali, Chinese (Mandarin and Cantonese), Dutch, Farsi, Filipino, French, German, Greek, Gujarati, Hebrew, Hindi, Indonesian, Italian, Japanese, Kannada, Korean, Malayalam, Marathi, Polish, Portuguese, Punjabi, Russian, Spanish, Tamil, Telugu, Thai, Turkish, Ukrainian, Urdu, Vietnamese.
There are more—over a hundred total—but those are the ones we've tested extensively and can confirm sound significantly better than before.
Why it took us this long
Honestly? Cost and complexity.
Running two different voice engines means more infrastructure. Language detection adds processing overhead. Routing logic needs to be bulletproof because if it fails, people get the wrong voice and the whole thing breaks.
We delayed this because we weren't sure we could do it without raising prices or making the system noticeably slower.
Turns out we could. Language detection is fast enough that it doesn't add perceptible latency. The routing happens automatically. And the cost difference balances out because English—which gets the cheaper engine—still makes up the majority of our traffic.
So we're shipping it now without a price increase.
What this does not fix
This improves quality for non-English languages dramatically, but it doesn't make everything perfect.
Highly technical jargon still trips up TTS in any language. Proper nouns get mispronounced sometimes. Poetry and lyrical writing lose nuance when spoken by a machine, regardless of the engine.
And if you're mixing languages in the same article—like writing primarily in English but quoting passages in Arabic—the system picks the dominant language and uses that voice throughout. It won't switch mid-article.
Those are limitations we're aware of. Some we can fix eventually. Some are just boundaries of what current TTS technology can do.
You don not need to do anything
That's the whole point of this update.
Send an article to @OutloudAIBot like you always do. The system detects the language, picks the right voice, and returns audio. Same workflow. Same speed. Better quality for most of the world.
If you've been converting non-English articles and tolerating mediocre audio, try sending one now. You'll hear the difference within the first ten seconds.
Why we are telling you this
Most companies would ship this silently and call it "continuous improvement." Just make things better without announcement and move on.
We're writing about it because the quality gap was real, and people noticed. If you'd been using OutloudAI for Farsi or Arabic content, you knew the audio wasn't great. You probably assumed that was just how TTS worked for your language.
It's not. It can be significantly better, and now it is.
This update matters most to users we'd been serving poorly. Seems worth telling them we fixed it.
Try sending an article in your language. If it sounds better than it did last week, that's why.