Voice cloning software that starts from ten seconds of audio

Updated October 28, 2026 · 8 min read

Most voice cloning tools are sold as part of a production suite: a timeline editor, a subscription, a render queue. The part you actually want is narrower than that. You want to record ten seconds of someone speaking, and then have that voice read text you type later — including sentences the person never said. That is what Resona does, in a browser, billed per use rather than per month.

This page describes what the cloning feature does, what it costs, what it will not do, and how to tell before you pay whether it is the right tool for your project.

What you get when you clone a voice

• A reusable model, not a one-off file. Cloning produces a voice you keep in your list. You do not re-upload the sample each time you want new audio — you pick the voice and type new text, months later, and it still works.
• One charge of 100 credits. Cloning is billed once per voice, roughly $2. It is not a recurring charge and it is not per generation.
• 83 languages from one sample. A voice cloned in English can read German, Japanese or Spanish text. See the accent caveat below — it matters more than most vendors admit.
• Emotion and pace controls on top. The clone sets the identity of the voice; style tags bias delivery toward calm, excited, urgent and so on without changing whose voice it is.
• No install and no GPU. Recording, cloning and generation all happen in the browser console. There is nothing to download and no model weights to manage.

How the cloning process works, step by step

1. Record or upload a sample

Anything from ten seconds to three minutes, WAV or MP3. Ten seconds is enough for the model to identify the voice. It is not enough to learn how the person phrases a sentence, which is the difference between a clone that is recognizable and a clone that is convincing.

2. The model extracts two separate things

Timbre is the colour of the voice — the part that makes it recognizable on the phone. Prosody is rhythm, stress and intonation. The first comes from a few seconds of audio. The second needs the speaker to talk naturally for longer, which is why a minute of unscripted speech beats a perfectly read paragraph.

3. The voice is stored in your library

Cloning finishes in minutes. The result appears alongside the 200,000+ public voices available without any sample at all — the two are interchangeable once the model exists.

4. Type text, choose the voice, generate

Speech generation is a separate line item, billed by volume. Long-form work is usually done chapter by chapter rather than as one enormous request, because it makes re-recording a single paragraph cheap when you find a mistake.

What it costs

Credits are prepaid and never expire. $1 buys 50 credits, packs start at $9, and there is no monthly fee — an idle month costs nothing, which matters if you narrate one project a quarter rather than one a week.

Speech is billed by UTF-8 byte rather than by character, which is the same unit the underlying engine uses. English is one byte per character; Chinese and Japanese are three. A 90,000-word novel is roughly 500,000 characters, so about 750,000 bytes and roughly 3,150 credits of speech — around $63 — after the $2 clone. Full line-item detail is on the pricing page.

What decides whether the clone sounds like the person

Four factors, in descending order of impact:

• Room noise beats microphone quality. A phone recording in a quiet room sounds better than a condenser microphone next to a running refrigerator. Fan noise, traffic and room reverb are what the model latches onto.
• Conversational speech beats read-aloud speech. Reading a paragraph flattens the pitch range. Answering a question does not.
• Length buys prosody. Sixty seconds of natural talking is worth more than ten seconds of studio audio, for everything except absolute clarity.
• Language coverage. Major languages are close to indistinguishable. Rarer ones still sound synthetic even with a good sample — that is a model limit, not a configuration mistake.

Our sample recording guide turns these into a checklist you can run before uploading.

What it does not do

Every vendor page lists capabilities. These are the limits, stated plainly, because finding them after you have paid is worse:

• Mannerisms are not cloned. Filler words, habitual pauses and verbal tics live in the script, not the voice. If the written text does not contain "um", the audio will not contain it either.
• Singing is out of scope. These are speech models. Pitch-accurate singing is a different problem with different tooling.
• Emotion is steerable, not automatic. Without a style tag or reference audio, output trends neutral. You direct the performance; it does not infer one from the text.
• Accent travels across languages. A voice with a strong regional accent will usually keep an audible trace of it when reading a second language. Sometimes that is exactly what you want; sometimes it is not.
• It will not fix a bad script. Cloning changes who is speaking, not what is being said.

Cloning compared with the alternatives

The honest framing: cloning is the right choice when identity matters (your own voice, a narrator whose rights you hold) and the volume is high. It is the wrong choice when you need a human to interpret a text — a memoir with irony, a novel with six distinct characters.

Who this fits

• Authors narrating their own book who cannot spend 40 hours in a booth, but whose voice is part of the product.
• Teams producing content at volume — course modules, product explainers, internal training — where per-word cost decides whether a project happens at all.
• Anyone localizing one script into several languages without losing the brand voice the audience recognizes.
• Not a fit for projects that need singing, simultaneous multi-character dialogue, or a performance the script does not describe.

Questions people ask before paying

Related: producing an audiobook with a cloned voice, text-to-speech generation in 83 languages, speech-to-text transcription, how neural cloning differs from older concatenative systems, and consent and legality, plus voice cloning rights and consent for dubbing.