ElevenLabs alternative: open-source text-to-speech and voice cloning on your own GPU

Local text-to-speech and on-premise voice cloning instead of ElevenLabs

Two open-source models can do a company's text-to-speech on its own workstation instead of ElevenLabs: Chatterbox Multilingual V3 from Resemble AI (MIT License) and Qwen3-TTS from the Qwen team (Apache 2.0) both speak German, English and many other European languages, clone a voice from a short reference recording and are licensed for commercial use. One NVIDIA RTX PRO card with 24 GB or more runs them, so scripts, voice recordings and finished audio stay in the company network. ElevenLabs still covers more languages, and in the only published German comparison that includes it, the Qwen team's own, it pronounced more clearly while Qwen3-TTS matched the reference voice more closely.

The typical jobs are e-learning narration, prompts for the phone system (IVR), voice tracks for product explainers and read-aloud functions for accessibility; the hardware is in one section with a table. Two related tasks have their own guides: AI video dubbing with lip sync for translating existing videos, and the voice agent server for an AI that answers calls live.

Text-to-speech (TTS) turns written text into spoken audio. Voice cloning makes a TTS model speak in a particular person's voice from a reference recording, often only a few seconds long. A model is the trained AI itself, a set of files loaded onto the graphics card (GPU); its licence says what a company may do with it and its output.

ElevenLabs languages, cloning rules and data residency

ElevenLabs' model overview, read on 30 October 2026, lists Eleven v4 with "90+ languages" and Eleven v3 with "70+ languages". Its voice cloning guide describes two kinds of clone. Instant Voice Cloning works from "1–2 minutes of good audio" and trains no custom model. Professional Voice Cloning trains a dedicated model on "30–180 minutes of good audio", and the guide sets a firm rule: "You can only create a Professional Voice Clone of your own voice", and "Even with their consent, you cannot clone someone else's voice."

The data residency page states that data is stored in the U.S. by default, that storage in the EU, India and Singapore is available, and that "Data residency is an Enterprise feature". It adds: "While storage will take place in the selected location, processing may nevertheless occur outside of the selected location." For EU residency, processing can be limited to the EU with Zero Retention Mode through the API.

With a local model the script, the reference recording and every generated file stay on the workstation, and the internet is needed only to download models and updates. In return, someone on the team tests the voices and looks after the machine.

Open TTS models and their licences, October 2026

GitHub repositories, Hugging Face model cards and licence files of resemble-ai/ResembleAI, QwenLM/Qwen, hexgrad, OHF-Voice and rhasspy, coqui, SWivid and fishaudio, read 30 October 2026. Company use as the licences read; your legal counsel confirms each use.

Chatterbox Multilingual V3 is Resemble AI's 0.5-billion-parameter model under the MIT License. Its 23 languages include German, English, French, Italian, Spanish, Polish, Dutch, Greek and the Scandinavian languages, but not Czech, Hungarian or Romanian. The card recommends it as "one voice cloning model that works across many languages". "Every audio file generated by Chatterbox includes Resemble AI's Perth (Perceptual Threshold) Watermarker", an inaudible mark that, according to the repository, survives MP3 compression and editing.

Qwen3-TTS was published on 22 January 2026 in sizes of 0.6 and 1.7 billion parameters, for Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian. The Base models offer a "3-second rapid voice clone from user audio input" and can be fine-tuned on one speaker. The CustomVoice models bring nine preset voices with control by written instruction, none of them a native German speaker, and the VoiceDesign model creates a new voice from a written description. Repository and model card state Apache 2.0; Hugging Face holds no separate licence file.

Kokoro-82M (Apache 2.0) has 54 preset voices in American and British English, Spanish, French, Hindi, Italian, Japanese, Mandarin and Brazilian Portuguese, but no German voice, and its card offers no voice cloning. For a European company it is an option for English narration.

Piper is a fast local engine. Its original repository was archived on 6 October 2025, and development moved to piper1-gpl under GPL-3.0. Its voice list states that "Piper is intended for personal use and text to speech research only" and that the licence of each voice is in its model card, so Piper is not part of the company set-up we offer.

Not for company use. XTTS-v2 offers "Voice cloning with just a 6-second audio clip" in 17 languages. Its licence file, the Coqui Public Model License 1.0.0, grants use "for any non-commercial purpose"; its summary says the licence "allows only non-commercial use of a machine learning model and its outputs", which covers the audio as well. F5-TTS releases its code under MIT, but "The pre-trained models are licensed under the CC-BY-NC license due to the training data Emilia". Fish Audio's OpenAudio S1-mini is under CC-BY-NC-SA-4.0. All three appear in lists of ElevenLabs alternatives, yet none of them may voice a company's videos, courses or phone prompts.

You can order the workstation alone, the setup of Chatterbox and Qwen3-TTS 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.

How well the open models speak German

The only German comparison with ElevenLabs we found comes from the Qwen team itself. On its model card, on what it calls the "TTS multilingual test set", Qwen3-TTS 1.7B Base reaches a German word error rate (WER) of 1.235 and a speaker similarity of 0.775. ElevenLabs reaches 0.572 and 0.614. The word error rate counts the words a speech recogniser gets wrong when it transcribes the generated audio, so lower means clearer pronunciation; speaker similarity measures how close the clone sounds to the reference voice. The card does not say which ElevenLabs model was tested, and the measurement is the vendor's own.

We found no published German listening test for any of these models.

A test with your own text settles it. Take 30 sentences from your material, with product names, numbers, dates, abbreviations, long compound nouns and English loanwords, and generate them with each model. A native speaker rates them, and Whisper transcribes the audio so that you can compare the transcript with the script, which is the same method as Qwen's word error rate. Our guide to German speech recognition shows how accurate Whisper is in German.

Consent, GDPR and AI Act rules for cloned voices

A cloned voice belongs to a person. Article 4(1) of the GDPR defines personal data as "any information relating to an identified or identifiable natural person", and a voice recording of an employee or a speaker falls under it. Agree in writing with the person whose voice is cloned on the purpose, the languages, the period of use and what happens to the voice model and the audio if the agreement ends. For employees the European Data Protection Board warns in its Guidelines 05/2020 on consent that consent is rarely free: "It is unlikely that an employee would be able to respond freely". With a hired speaker, the use of the voice can be settled in the contract for the recordings. Your data-protection officer confirms the legal basis, and our guide on local AI and the GDPR explains what own hosting changes.

The AI Act adds two duties. Under Article 50(2), providers of AI systems that generate synthetic audio ensure that "the outputs of the AI system are marked in a machine-readable format" and detectable as artificially generated. The AI Act counts as a provider anyone who develops an AI system, or has one developed, and puts it into service under its own name, and putting into service includes supply "for own use", so a company that has its own voice system set up can fall under this; our article on the EU AI Act for companies explains the roles. Article 50 applies from 2 August 2026, and providers of systems placed on the market before that date have until 2 December 2026 under Article 111(4). Chatterbox embeds an inaudible watermark in every file, and its repository provides a script to detect it; whether that meets the requirement is for your legal counsel to judge. The Qwen3-TTS and Kokoro pages mention no watermark; Resemble AI also publishes its watermarker, Perth, as a separate library under the MIT License that marks any audio file, so the output of Qwen3-TTS can carry the same mark.

Under Article 50(4), deployers disclose a deep fake, which Article 3(60) defines as content that "resembles existing persons" and "would falsely appear to a person to be authentic or truthful". A clone of your managing director reading a text she never read can meet that definition; a voice designed from a description, which resembles no one, is a different case. Your legal counsel confirms what each use needs.

Which GPU for local text-to-speech and voice cloning

Text-to-speech models are small. The repositories publish no GPU memory figures, so the table works from the size of the model files and from OpenAI's figure of about 10 GB for Whisper large, which checks the output. By our arithmetic, Chatterbox Multilingual V3 takes about 3.2 GB (its v3 files of 2.14 and 1.06 GB), Qwen3-TTS 1.7B Base about 4.5 GB (3.86 GB plus its 0.68 GB speech tokenizer), and the three together about 17.8 GB, or 16.5 GiB. 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, 40.0 on 48 GB, 61.5 on 72 GB and 83.0 on 96 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; working memory comes on top of the files.

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 long texts are better generated one model at a time. The RTX PRO 4500 Blackwell with 32 GB at 200 W leaves about 9 GiB and is the more comfortable choice for daily work.

Team. The RTX PRO 5000 Blackwell has 48 or 72 GB at 1,344 GB/s and draws 300 W. It adds a local translation model, for example EuroLLM-9B-Instruct (Apache 2.0, about 18.3 GB of weights in 16-bit), so that a German script becomes French, Polish or Spanish narration on the same machine: about 36.1 GB, or 33.6 GiB, with the voice models and Whisper, within 40.0 GiB on the 48 GB card. Qwen's fine-tuning guide for a single speaker gives no memory figure, so a trial run on your recordings shows what it needs; the 72 GB card leaves the most room. Our guide to on-premise machine translation covers the translation models.

Mid. One RTX PRO 6000 Blackwell Workstation Edition or Max-Q with 96 GB renders several voices in parallel and keeps fine-tuning and production apart. The Workstation Edition draws 600 W, the Max-Q 300 W with a blower cooler. Two Workstation Edition cards are the realistic maximum for that edition.

Large. Read-aloud on demand for a website, or a phone system with many lines, needs real-time speech for many listeners at once. That is server work, sized in our guide to the voice agent server.

Speed is not published for RTX PRO cards. Qwen states a synthesis latency "as low as 97ms" without naming the hardware, and Resemble AI states that its small English model Chatterbox-Nano runs "3x faster than realtime on 8 CPU cores". How long an hour of narration takes depends on the model, the text length and the settings; have one of your usual scripts timed on the delivered workstation.

Setting it up for courses, phone prompts and read-aloud

E-learning narration. Courses are produced in batches: one script, many short audio files. Qwen3-TTS includes a browser demo started with qwen-tts-demo, and both models can be called from a short script that turns a folder of text files into audio files. A fixed reference recording per voice and a pronunciation list for product names keep the voice consistent across modules, as our guide to AI voice-over for marketing videos describes, together with timing to the edit.

Phone prompts. The phone system plays back short prompts, and the audio must be converted to the format it expects, for example G.711 with 8,000 samples per second, as our contact-centre guide notes. Generate the prompts at the model's own sample rate, Kokoro's examples use 24,000 Hz, and convert afterwards.

Read-aloud for accessibility. Prepared audio files need no live server, and text that changes daily is generated overnight.

For the IT team. RTX PRO Blackwell cards need CUDA 12.8 or later, as our guide to Blackwell software support explains. Qwen's README starts its demo with --ip 0.0.0.0, which opens it to the whole network; keep such interfaces behind the company firewall with logins. Store reference recordings and voice models with the same access rules as personnel files.

Send us the languages, the number of voices and the minutes of audio per month, and we propose the card and the configuration that fit.

FAQ