AI voice-over for marketing videos, made locally: timed narration in several languages

An AI voice-over generator for marketing videos on your own workstation

A marketing team can generate the voice-over for product and explainer videos on its own workstation: the script is split into one line per scene, an open voice model such as Chatterbox Multilingual V3 or Qwen3-TTS speaks each line in a fixed brand voice, and the audio files go into the editing software like any recorded narration. Both models speak German, English and many other European languages and are licensed for commercial use. One NVIDIA RTX PRO 4500 Blackwell with 32 GB runs them, and scripts with unreleased product names stay in the company.

This guide is for marketing managers, video editors and agencies who make explainer, product and social videos in several languages and want the narration without a cloud voice service. Our comparison of local alternatives to ElevenLabs covers the voice models, their German quality and voice cloning in general; here the subject is the production workflow. For translating existing videos with lip sync, see our guide to AI video dubbing.

A voice model, or text-to-speech (TTS) model, turns text into spoken audio. A take is one generated version of a line; voice models vary slightly from take to take, as a speaker does. An SRT file is a plain subtitle file in which each line of text has a start and an end time.

From script to voice track in seven steps

1. Write for the ear. Short sentences, one idea each. Numbers, units and abbreviations are written the way they should be spoken.
2. Split by scene. One line per shot or scene, numbered, so that the audio files arrive in the order of the edit. If the timing is already fixed, write the lines into an SRT file with the start time of each scene.
3. Choose the voice. A preset voice, a designed brand voice or the cloned voice of a person who agreed. The next section explains the three.
4. Fix pronunciation. Brand and product names go into a pronunciation list before the first take.
5. Generate takes. Two or three takes per line, then pick the best by ear.
6. Check. Whisper, OpenAI's speech recognition model under the MIT License, transcribes the finished track; any word in the transcript that differs from the script marks a line to listen to again. The same run writes the subtitles: Whisper's command-line tool offers srt among its output formats.
7. Edit. The WAV files go onto the timeline in the editing software, where music, mix and loudness are done as with any narration.

On the workstation the steps run in ComfyUI, an interface under GPL-3.0 in which a prepared workflow is started with a few clicks, or in the browser demos that the voice models bring. Whoever writes the script and whoever edits the video can work on the same machine.

Choosing a preset, designed or cloned voice

Preset voices. Kokoro, a small model under Apache 2.0, ships 20 American and 8 British English voices, plus voices in Spanish, French, Italian, Hindi, Japanese, Mandarin and Brazilian Portuguese. It has no German voice. For English videos it is the quickest start.

A designed brand voice. Qwen3-TTS from the Qwen team, released on 22 January 2026 under Apache 2.0, includes a VoiceDesign model that "Performs voice design based on user-provided descriptions": you describe the voice in a few sentences, its age, pitch and character, and the model creates it. Its README then describes "a practical workflow": use VoiceDesign "to synthesize a short reference clip that matches your target persona" and give that clip to the Base model, which clones from three seconds of audio. The README calls this "especially useful when you want a consistent character voice across many lines". For marketing this is the most useful path: the voice belongs to the brand and resembles no real person. Qwen3-TTS speaks German, English, French, Italian, Spanish, Portuguese and Russian among its 10 languages; its nine preset voices are native Chinese, English, Japanese and Korean speakers who, by its README, can speak every supported language; for a German brand voice, test a designed voice first.

A cloned voice. Chatterbox Multilingual V3 from Resemble AI, under the MIT License, and the Qwen3-TTS Base models clone a person's voice from a short recording, for example the voice of your product manager or of a speaker you have hired. That needs a written agreement on purpose, languages and duration; our ElevenLabs comparison explains the consent questions.

Keep the reference clip in a fixed place and use it for every video. Chatterbox adds two controls for delivery: exaggeration, default 0.5, which can go to about 0.7 for more expressive speech, and the CFG weight, default 0.5. Its card notes that higher exaggeration tends to speed up speech and that a lower CFG weight helps slow it down.

You can order the workstation alone, the setup of the voice models and ComfyUI alone, or both together: the workstation delivered with everything installed and configured; write to us and we suggest the best solution for your tasks and budget.

Product names, numbers and pronunciation

AI voices often stumble over brand and product names, because a new name does not appear in the material the model was trained on. Three methods help, in this order.

Respell. A pronunciation list maps each name to a spelling the model reads correctly, for example a name with an unusual stress written with a hyphen or a vowel doubled. The list is applied to every script before generation and kept with the brand guidelines.

Phonetic spelling. Kokoro accepts the pronunciation inline in the text, in the form [Kokoro](/kˈOkəɹO/) from its README, where the part in slashes is the pronunciation in phonetic symbols. TTS Audio Suite, a set of ComfyUI nodes under the MIT License, includes a Phoneme Text Normalizer with "Full International Phonetic Alphabet conversion using espeak"; whether a given voice model benefits from it is worth a test.

Test sentences. Put every product name into a short test sentence and generate it with each voice before a campaign starts. Numbers, prices spoken in a currency, dates and units belong on the same list, because voice models read them differently. The Whisper check from step 6 catches names that come out wrong.

Timing the voice-over to the edit

The simplest method is one audio file per scene: the editor places each file at the start of its scene and moves it by hand. When a line is too long for its shot, shorten the text and generate it again, which sounds better than stretching the audio.

From an SRT file. TTS Audio Suite has an SRT node that generates speech from a subtitle file and aligns the audio to its timings. Its smart_natural mode uses "Intelligent shifting logic that prevents overlaps and ensures natural speech flow", and an Adjusted_SRT output reports the timings of the generated audio, so you see which lines ran longer than planned. The suite supports Chatterbox Multilingual and Qwen3-TTS among many other engines; the README listed version 5.8.11 when we read it on 30 October 2026.

Speed. Kokoro has a speed setting (speed=1 in its example). Chatterbox's documentation names no speed setting; its pace follows the exaggeration and CFG weight described above. The CustomVoice and VoiceDesign models of Qwen3-TTS take written instructions that, according to the model card, give "flexible control over timbre, emotion, and prosody".

Several voices in one video are possible too: TTS Audio Suite switches voices with tags such as [Alice] at the start of a line, and languages with tags such as [de:Alice].

Language versions from one script

For a campaign in German, English, French and Polish, the master script is translated first, by a translator or a local language model, as in our guide to machine translation on-premise. A native speaker of each language reads the translation before any audio is made and shortens lines that run longer than the scene.

Resemble AI's card recommends Chatterbox Multilingual V3 as "one voice cloning model that works across many languages", so the same brand voice can speak all versions. The card also warns that the reference clip should match the language of the text, otherwise the output may take on the accent of the clip's language; setting the CFG weight to 0 reduces this. For a brand voice in several languages, record or design one reference clip per language, and have each version checked by a native listener.

Licences and AI Act labelling for advertising

GitHub repositories and Hugging Face cards of Resemble AI, Qwen, Kokoro, TTS Audio Suite and OpenAI, read 30 October 2026; ComfyUI licence as recorded in our AI video workstation guide.

Videos for customers are commercial use, and each model's licence must allow it. The tools in the table do. TTS Audio Suite also runs engines with other terms: its licence file lists the licence of each engine's weights and marks several, F5-TTS among them, as not for commercial use and others as conditional, so the workflow should load only approved engines.

The AI Act has two rules for synthetic voices, both applying from 2 August 2026. Article 50(2) requires providers of systems that generate synthetic audio to mark it "in a machine-readable format"; a company that sets up such a system for its own use can count as a provider, as our article on the EU AI Act for companies explains. Chatterbox puts Resemble AI's inaudible Perth watermark into every file; the Qwen3-TTS and Kokoro pages mention no watermark.

Article 50(4) requires deployers, the companies that use such a system, to disclose a deep fake, which Article 3(60) defines as AI-generated or manipulated image, audio or video content that "resembles existing persons" and "would falsely appear to a person to be authentic or truthful". In an advert the choice of voice matters. A designed brand voice is not modelled on a real person, while a preset voice may be based on recordings of a real speaker; the cloned voice of your product manager presenting a product in words she never recorded can meet the definition. Article 50(5) asks for this information in a clear and distinguishable manner at the latest at the first exposure, so a note visible from the start of the video, or a spoken line at its beginning, is one way to disclose it; the inaudible watermark tells the viewer nothing. Your legal counsel confirms what each campaign needs.

GPU memory for voice models and the card to choose

Voice models are small. The repositories publish no GPU memory figures, so we work from the size of the model files: about 3.2 GB for Chatterbox Multilingual V3, about 4.5 GB for Qwen3-TTS 1.7B Base with its speech tokenizer, and about 10 GB for Whisper large by OpenAI's figure. Together that is about 17.8 GB, or 16.5 GiB, by our arithmetic. Our sizing rule counts 90 per cent of the memory the driver reports, less 3 GiB, as usable: 18.5 GiB on a 24 GB card, 25.7 on 32 GB and 40.0 on 48 GB.

file sizes from the Hugging Face repositories of ResembleAI and Qwen and OpenAI's Whisper README, read 30 October 2026; usable memory by our sizing rule; card data from NVIDIA's datasheets. Our arithmetic, not measurements.

Entry. The NVIDIA RTX PRO 4000 Blackwell has 24 GB at 145 W in one slot. The three models fit within its 18.5 GiB with about 2 GiB to spare, so longer scripts are better generated with one model loaded at a time.

Recommended. The RTX PRO 4500 Blackwell has 32 GB with 896 GB/s at 200 W. It leaves about 9 GiB after the three models, for longer lines, several takes and the editing software on the same machine. This is the card we suggest for a marketing team. Both cards run from an ordinary office socket.

Step up. The RTX PRO 5000 Blackwell with 48 or 72 GB at 300 W keeps a translation model such as EuroLLM-9B-Instruct (about 18.3 GB in 16-bit) loaded next to the voice models: about 33.6 GiB in total by our arithmetic, within the 40.0 GiB of the 48 GB card. It also suits videos in which an AI avatar speaks, as our guide to local AI avatar videos describes.

Mid. When the same machine also generates the footage, the video models set the size: our guide to the AI video workstation for marketing recommends one RTX PRO 6000 Blackwell Workstation Edition or Max-Q with 96 GB for the large open video models, and the voice models run on the same card between video jobs.

No publisher states how long a minute of narration takes on an RTX PRO card. It depends on the model, the length of the lines and the number of takes; have one of your scripts timed on the delivered workstation before you plan a campaign.

Send us your languages, the number of videos per month and your editing software, and we propose the configuration that fits.

FAQ