Free voice AI: 10 options and the PC each one requires

You recorded a video, wrote the script and now need narration. ElevenLabs appears in every tutorial, but the plan that allows commercial use is paid. Is there free voice AI that can handle the job in Portuguese?

There is. Some run in the cloud and give credits every month; others are open source: you download them for free and run them on your computer, without limits. This guide lists 10, with the free limit, the paid plan and how much graphics card each one requires.

Quick summary: the answer in 30 seconds. To start without installing anything, use ElevenLabs, Gemini TTS, Azure Speech or Fish Audio, which give free credits every month. For unlimited use, install an open model: Kokoro and Piper run even without a graphics card; Chatterbox, Qwen3-TTS, VoxCPM2 and Voxtral clone voices; the last three require a graphics card with 6 to 16 GB of memory.

Index

• Cloud or on your PC: what's the difference between free voice AI and open source?
• 4 free voice AIs in the cloud, with monthly credits
• 6 open source voice AIs to run on your computer
• What PC does each voice AI need?
• Which one to choose?
• Be careful when cloning voices
• Frequently asked questions

Cloud or on your PC: what's the difference between free voice AI and open source?

In the cloud, the voice is generated on the company's servers: you just need a browser, but the free tier has a monthly quota. With open source, the AI runs on your computer, without a quota. The price becomes hardware: the larger the model, the more graphics card it requires.

All do TTS (text-to-speech with AI): you type the sentence and receive the audio. The best ones also clone: they listen to a few seconds of your voice and start speaking with it.

To know if a model fits on your PC, look at two numbers. The first is the parameters, the "adjustment knobs" that the AI learned during training. Think of a translator's vocabulary: the larger it is, the more nuance, and the more space it occupies.

The second is VRAM, the graphics card memory, which holds the model while it works. It's the carpenter's workbench: if the piece doesn't fit, the work doesn't progress. If GPU and graphics card still confuse you, see the guide about GPU and graphics card.

The practical rule: billions of parameters × 2 = gigabytes of VRAM just for the model. A 2 billion parameter model occupies about 4 GB, plus 2 or 3 GB of headroom for the work. It's an estimate, but it gets the order of magnitude right.

4 free voice AIs in the cloud, with monthly credits

All four have a free plan renewed every month and work in the browser: any laptop or phone with internet will do. They differ in quota size and what the free tier allows, such as voice cloning or commercial use.

1. ElevenLabs

ElevenLabs is the market reference, valued at $11 billion in February 2026, according to TechCrunch.

The free plan gives 10,000 credits per month, equivalent to about 10 minutes of speech. The commercial license and instant voice cloning only appear starting with the Starter plan, at $6 per month with 30,000 credits (checked on 10/27/2026).

Portuguese is well served: version v3 of the voice model covers more than 70 languages, and v4, more than 90.

2. Gemini TTS, from Google

The Gemini TTS documentation lists 30 ready-made voices, Portuguese among the languages and up to two voices conversing in the same audio, good for simulating a podcast.

You can test it in Google AI Studio without programming. For programmers, the Gemini 3.8 Flash TTS and Flash-Lite TTS models have a free quota; above it, billing is per use (checked on 10/27/2026). One advantage: you adjust the pace and tone with a text instruction, like "speaking slowly". Google also replicates voices, but requires a consent recording from the person.

3. Azure Speech, from Microsoft

Azure Speech is the most generous. The free F0 plan gives 500,000 characters per month in AI-generated voices (checked on 10/27/2026). At an average of about 900 characters per minute of speech, that yields about 9 hours of audio.

There are more than 20 voices in Brazilian Portuguese, such as Francisca, Antonio and Thalita, including multilingual versions. The price is bureaucracy: an Azure account and access key, a path designed for programmers.

4. Fish Audio

Fish Audio has more than 8 million users, according to TechCrunch. The free plan gives 8,000 credits per month, up to 7 minutes of generation and voice cloning included.

The catch: audio from the free plan is only for personal projects. Plus costs $11 per month (checked on 10/27/2026). The company also released for download the S2 Pro model, with 5 billion parameters and more than 80 languages, but the license is for research only.

In one sentence: for volume, Azure; for quality, ElevenLabs; to clone for free, Fish Audio.

6 open source voice AIs to run on your computer

Open source voice AI is a model you download for free and run on your PC, without a quota and without sending the text to anyone. In exchange, it requires installing Python (the language these programs run in) and following the step-by-step on GitHub, the site where the code is published. All six speak Portuguese; the first two don't require a graphics card.

5. Kokoro

Kokoro is the lightweight of the list: 82 million parameters, Apache 2.0 license (can be used commercially) and 54 voices. Among them, three in Brazilian Portuguese: Dora, Alex and Santa.

Hardware: by the rule, the model occupies less than 0.2 GB. It runs on any computer, just with the processor, and on Mac with Apple chip. The limitation: it doesn't clone voices.

6. Piper

Piper is a lightweight and fast voice engine. It speaks about 40 languages, including Brazilian Portuguese, and has a GPL-3.0 license (allows commercial use, but anyone who modifies the code must open it).

Hardware: the project itself distributes versions for Raspberry Pi 3 and 4, a card-sized computer. Graphics card is optional. The voice is more robotic than the others, and there's no cloning.

7. Chatterbox

Chatterbox, from Resemble AI, is the entry point for cloning voices at home. Multilingual V3, with 500 million parameters, covers at least 23 languages, including Portuguese, and there's a package tuned just for Brazilian Portuguese. MIT license, free for commercial use.

Hardware: by the rule, about 1 GB for the model, about 4 GB with headroom. It runs on NVIDIA graphics cards, on Mac with Apple chip or, more slowly, on the processor. All audio comes out with an inaudible watermark, PerTh, which identifies the voice as AI-generated.

8. Qwen3-TTS

Qwen3-TTS, from Alibaba, came out on January 22, 2026 in two versions: 0.6 and 1.7 billion parameters. It speaks 10 languages, Portuguese among them, and clones a voice with 3 seconds of audio. Apache 2.0 license.

Hardware: by the rule, from 1.2 to 3.4 GB just for the model. In practice, the larger version requires an NVIDIA card starting at 6 GB; an RTX 3060 with 12 GB runs comfortably.

9. VoxCPM2

VoxCPM2, from OpenBMB, has 2 billion parameters, 30 languages (Portuguese included) and generates audio at 48 kHz, professional video standard. It clones voices with a short clip and has an Apache 2.0 license.

Hardware: the number is official: about 8 GB of VRAM, on an NVIDIA card. On an RTX 4090, a top-of-the-line gaming card, it generates one minute of audio in about 18 seconds.

10. Voxtral TTS, from Mistral

Voxtral TTS, launched by French company Mistral on March 23, 2026, has 4 billion parameters and 9 languages, Portuguese included. It clones a voice with 3 seconds of reference and can speak another language while maintaining the cloned timbre. See the context in the text about Voxtral 3.

Hardware: the model card requires a card with 16 GB or more, like the RTX 4080. The license is CC BY-NC, non-commercial; to sell the audio, use the API, at $0.016 per thousand characters (checked on 10/27/2026), less than $1 per hour of audio.

What PC does each voice AI need?

The cloud only requires a browser. Kokoro and Piper run on any computer, without a graphics card. Chatterbox and Qwen3-TTS require an entry-level card, 4 to 6 GB; VoxCPM2 wants 8 GB, and Voxtral, 16 GB. Values with "~" are estimates by the rule.

Which one to choose?

It depends on the computer and the use: a regular laptop stays with the cloud or Kokoro; those who have a graphics card and want to clone voices go with Chatterbox or Qwen3-TTS. To sell the audio, the license weighs more than the quality.

Be careful when cloning voices

Cloning a voice is only acceptable with the authorization of the speaker. Scammers use cloned voices to ask relatives for money over the phone, and several tools on this list clone with seconds of audio.

Fish Audio faced accusations from creators whose voices were uploaded by third parties without consent, according to TechCrunch. Watermarks like Chatterbox's help track the audio, but don't prevent misuse.

And free services can disappear: Hume will shut down its voice API on November 13, 2026. Don't depend on a single tool.

Frequently asked questions

Can I use ElevenLabs' free plan in a monetized video?

No. The commercial license only appears starting with the Starter plan, at $6 per month. To monetize without paying, use an open model with an Apache or MIT license, such as Kokoro, Chatterbox or Qwen3-TTS.

Which voice AI sounds most natural in Portuguese?

There's no official ranking in Portuguese; your ear decides. Azure has the largest variety of Brazilian voices, and Chatterbox has a package tuned for Brazilian Portuguese. Test the same text in two or three.

Can I clone my own voice for free?

You can. In the cloud, Fish Audio's free plan clones, only for personal use. On your PC, Chatterbox and Qwen3-TTS clone without limits and allow commercial use.

Do I need to know how to program to use open source models?

A little. All six require installing Python and running commands in the terminal. Some provide a demo page in the browser: after installation, just type the text.

To start today

Test a cloud one today and, if you like it, install Kokoro: it runs on the computer you already have and costs nothing. To follow the sector, check the artificial intelligence section.