Text to speech that reads your script the way punctuation intends it
Updated October 29, 2026 · 7 min read
Long articles, newsletters, manuals, course modules. Turning them into audio used to mean booking a narrator or installing a synthesis stack. Resona does it in a browser: paste the text, pick a voice, generate. Speech comes out in MP3 or WAV, in any of 83 languages, and you pay per generation rather than per month.
This page covers what the feature does, what it costs, what decides quality, and what it will not do — including the limits most vendor pages leave out.
Text to speech and voice cloning are different jobs
The two get confused constantly, and the confusion leads people to pay for the wrong thing:
- Text to speech (text → audio). You have a script and want audio. Pick a stock voice from the 200,000+ public library, or pick a voice you cloned earlier. You do not need to create anything new to start.
- Voice cloning (voice → model). You want a specific person's voice — usually your own — available for generation later. Cloning is a separate one-time step: 100 credits per voice, and the model stays in your list permanently.
If you just need an article read aloud, start with a stock voice. Cloning is for when the identity of the voice is part of the product.
How it works
1. Paste the text
Punctuation and line breaks become pauses and intonation, so writing for the ear is the same skill as reading well. Long scripts can go in one request, but splitting by section makes later fixes cheap: regenerate one paragraph instead of one chapter.
2. Pick the voice
Choose from the public voice library or from voices you cloned yourself. Both work in the same console. Output languages include Japanese, German, Spanish and 80 others — the model is language-independent, so one voice can read across languages.
3. Generate, listen, download
Audio plays in the browser and downloads as MP3 or WAV. Before shipping, check it on the device your audience will use — phone speaker, car, single earbud. Room acoustics hide flaws that headphones do not.
What it is used for
- Audio versions of written content. Articles and newsletters reach readers who prefer listening, from the same script.
- Video and course narration. One consistent voice across dozens of modules, without scheduling a narrator for each revision.
- Language learning. Hear any text pronounced on demand; turn study notes into audio for a commute.
- Accessibility. Text-only pages gain an audio path for people who cannot or will not read on screen.
- Documentation and manuals. Hands-busy audiences can listen to procedures instead of switching between page and task.
What it costs
Generation is billed by UTF-8 bytes, not characters: 4,200 credits per 1,000,000 bytes. One credit is $0.02 when you buy a pack ($9 buys 450), and $1 buys 50 credits. Credits are prepaid, never expire, and there is no monthly fee.
| Amount | Bytes | Credits | Approx. cost |
|---|---|---|---|
| 1,000 English characters | ≈ 1,000 | ≈ 4.2 | ≈ $0.08 |
| 10,000 English characters | ≈ 10,000 | ≈ 42 | ≈ $0.84 |
| 10,000 Japanese characters (3 bytes/char) | ≈ 30,000 | ≈ 126 | ≈ $2.52 |
| 100,000 Japanese characters | ≈ 300,000 | ≈ 1,260 | ≈ $25.20 |
The byte distinction matters: Japanese and Chinese are three bytes per character, so identical character counts cost about three times more than English. A one-hour English narration is usually a few dollars. Using a cloned voice adds the one-time 100-credit clone charge — generation itself prices the same for stock and cloned voices. The full rate table is on the pricing page.
What decides quality
- Punctuation is the prosody. Commas take a breath, periods fall, exclamation marks bite. Flat-sounding output is usually fixed by editing punctuation, not by switching voices.
- Request length. Section-by-section generation lets you set a different tone per section and keeps mistakes cheap to fix.
- Language and pronunciation. In-language text is stable. Proper nouns, jargon and abbreviations deserve a human final check — that is where errors concentrate.
- Emotion tags where supported. Some models accept style tags that bias delivery toward calm, excited or urgent without changing the voice. The console shows which models accept tags.
What it does not do
- Multi-character performance. One voice is one persona. Fiction with distinct characters means one voice per character and line-by-line assembly — production work, closer to directing than narrating.
- Comedic timing from context. The generator reads the chunk in front of it and does not know what comes next. The pause that lands a joke is a human narrator's contribution.
- Perfect proper nouns. Rare names and acronyms produce the kind of errors a listen-through catches. Always do a final pass before publishing.
- Singing. These are speech models; pitch-accurate singing is a different problem.
Privacy
Input text is sent to our technology partner for processing only. It is not sold and not used to train models. What is stored, for how long, and your rights are in the privacy policy. If the voice you use is a clone of a real person, holding that person's consent is your responsibility — see the consent and legality guide.
Questions people ask before paying
What audio formats can I download?
MP3 and WAV. MP3 for video narration and publishing; WAV when the audio will be edited further.
How is speech generation billed?
By UTF-8 bytes of input text: 4,200 credits per 1,000,000 bytes. One credit is $0.02 in a pack; $1 buys 50. English is about one byte per character, Chinese and Japanese three — so the same character count in Japanese costs roughly three times more. A cloned voice adds a one-time 100-credit charge for the clone itself.
Can I use my own cloned voice?
Yes. Clone once (100 credits) and the voice sits in your list next to the 200,000+ public voices, usable in the same generation flow.
Why does Japanese text cost more than English of the same length?
Billing is by UTF-8 byte. Japanese is three bytes per character, English one — ten thousand Japanese characters are about 30,000 bytes against about 10,000 for English.
Which languages are supported?
83 languages for generation. The voice model is language-independent: one cloned voice can read across languages, with the accent following the training audio.
Related: voice cloning from a 10-second sample, speech-to-text transcription, and what AI voiceover costs per minute.
日本語版: テキスト読み上げツール — 音声合成の使い方と料金.