12 November 2026 · VRAMfarm team

Text to speech on your own GPU: Qwen3-TTS, VoxCPM2 and Chatterbox voice cloning

A 40-minute training course needs a narrator, in English and German, by Friday. On a rented 12 GB card with an open-source voice model, the voice-over takes an afternoon and roughly €0.18 of GPU time on an RTX 3060. Text to speech is one of the few AI jobs where the cheapest cards in the catalogue are enough: the models are small, and the hard part is your script and your ears.

We have three ready-made speech apps, free to use; you pay only for GPU time, from €0.12 an hour on an RTX 3060. Each one is good at something different.

Three apps, three jobs

Qwen3-TTS is the simplest: choose one of the built-in voices, paste text, and add an instruction such as “cheerful, a little faster” if you want. Good for notifications, explainer videos and anything where a consistent stock voice is fine. The Qwen3-TTS family also has a separate cloning model, which you can install yourself on an empty server.

VoxCPM2 sounds the most alive of the three: pauses, questions and exclamations come out the way a person would read them. Use it for audiobooks, podcasts and long narration. It can design a voice from a description, or copy one from a short sample if you also give it the sample’s transcript.

Chatterbox is the one for cloning: upload a clean recording of a few seconds and it reads any text in that voice, in 23 languages. It also has an “exaggeration” control that makes the delivery flatter or more expressive.

All three open in your browser. The text-to-speech page has samples of the same text in each, which is the fastest way to choose.

VRAM and which card to rent

These are small models, a few gigabytes of weights each (VoxCPM2’s are about 5 GB). In use they take roughly 6–8 GB of VRAM, so a 12 GB card has enough room and our apps ask for 12 GB. The RTX 3060 is the cheapest that qualifies; an RTX 3080 with 10 GB is too tight for our apps.

A faster card does help, but less than you might expect. Speech models generate audio piece by piece, so a card two or three times faster in raw compute makes each sentence come back faster, not proportionally cheaper. Rough rule:

• RTX 3060 (12 GB): fine for a voice-over you listen to sentence by sentence.
• RTX 4090 or RTX 5090: worth it when you voice hours of text in a batch, or when waiting 20 seconds per paragraph slows you down.
• Anything with 48 GB or more: no benefit for these models. You pay for memory that sits empty.

Speed and cost per hour of narration

On a 12 GB card, a couple of sentences take somewhere between 10 and 30 seconds with Qwen3-TTS, and a single phrase 5–20 seconds with Chatterbox, depending on length. That works out to roughly 1–2.5 seconds of GPU time per second of speech on an RTX 3060, and about 0.4–1 second on an RTX 4090. These are rough planning ranges, not benchmarks: text length, settings and the model version move them, and the first sentence of a session is always slower while the model loads.

For an hour of finished narration (about 9,000 words, at the middle of each range):

Add the minutes you spend listening and redoing sentences, usually 10–20% of them, and a few minutes at the start while the app installs its packages and downloads the weights. Those install minutes are billed, as is the download traffic at the machine’s rate. Even with retakes and the install, an hour of narration on an RTX 3060 comes to roughly €0.26. The billing guide explains the 15-minute minimum per launch and why it makes sense to voice everything in one session.

Cloning a voice: the recording decides the result

With Chatterbox (or VoxCPM2), the quality of the clone depends far more on the sample than on the card.

• Length: 5–10 seconds of continuous speech. Longer samples help only if they stay clean.
• Clean: one speaker, no music, no room echo, no noise reduction artefacts. A phone held close in a furnished room beats a good microphone in a bare kitchen.
• The right tone: the clone copies the delivery too. Read the sample the way you want the narration to sound, calm for a course, lively for an ad.
• The right language: a sample in the target language gives the best accent. In Chatterbox, if the sample is in another language, the model’s guidance setting (CFG weight) at or near 0 reduces the accent carried over.
• For VoxCPM2: type the exact words of the sample into the transcript field. A wrong transcript makes the clone worse.

If the voice sounds rushed, lower Chatterbox’s CFG weight a little; if it sounds monotonous, raise the exaggeration. Change one setting at a time and keep the seed fixed while you compare.

Long texts without losing your mind

The apps are made for paragraphs, not whole books in one go. A workflow that holds up:

1. Split the script into paragraphs of two to five sentences. Redoing one paragraph is quicker than redoing a chapter.
2. Write numbers, dates and abbreviations as words. “Ten thousand euros” and “the twelfth of March” come out right; “€10,000” and “12/03” sometimes don’t.
3. Use punctuation to direct. Commas give short pauses, full stops longer ones, question marks lift the end.
4. Number your files (01.wav, 02.wav) as you download them from the player.
5. Join them afterwards in Audacity, your video editor or with ffmpeg.

For a long batch, the “Use via API” link at the bottom of each app shows how to call it from a script with Python’s gradio_client, one paragraph per call. That turns a 9,000-word script into a loop you start and come back to.

Download each clip as it is made. The apps keep generated audio in temporary files outside /workspace, so it is not in the saved folder and is gone when the server stops. Your reference recording and scripts can live in /workspace and survive Save and stop; the model weights are downloaded again on the next launch. See what happens to your files when you stop a GPU.

Whose voice is it? Consent and the EU AI Act

Cloning a voice is technically trivial now, which is exactly why the rules matter. This is general information, not legal advice; for a commercial project with someone else’s voice, ask a lawyer in your country.

• Your own voice: fine to clone and use.
• Someone else’s voice: only with their clear permission, ideally written, saying what it will be used for and for how long. That includes friends, colleagues and voice actors. Many voice actors’ contracts now mention AI cloning explicitly; check before you assume.
• Public figures and the deceased: don’t. Imitating a real person’s voice without consent can break personality, image or publicity rights in many countries, and in some it is a criminal matter when used to deceive.
• The recording is personal data. A voice sample of an identifiable person falls under GDPR, and a voice used to recognise someone counts as biometric data. Keep the sample only as long as you need it.

The EU AI Act adds transparency duties in Article 50, which start to apply from August 2026; some details and transition periods were still being adjusted during 2026, so check the current state before a commercial release. Broadly: systems that generate synthetic audio should mark it as machine-made, and if you publish audio that imitates a real person so that it could pass as authentic, you are expected to disclose that it is artificial. Lighter rules apply to clearly artistic or satirical work. Which duties fall on you depends on your role and use. A simple habit covers most cases: say in the credits or description that the voice is AI-generated.

Chatterbox adds Resemble AI’s inaudible watermark to every file it makes, which detection tools can read. As far as we know, Qwen3-TTS and VoxCPM2 don’t add one by default, so the disclosure is up to you.

Do you need your own GPU for this?

Honestly, not always. For a few sentences a week, a hosted TTS service or a small model on your laptop’s CPU is less effort, and some commercial services have more polished stock voices in certain languages. A rented GPU is the better deal when you narrate hours of text, need several languages from one voice, want a cloned voice that never leaves a machine you control, or don’t want a per-character bill. Start an RTX 3060 with the app that fits, voice the first paragraph, and you’ll know within 15 minutes whether it sounds right.