ChatGPT answers, unedited

If privacy is the priority, I’d look at local/offline TTS models rather than creator tools that send scripts to a cloud API.

Best options

1. Kokoro — my first recommendation.
Kokoro is an 82M-parameter open-weight TTS model that can run entirely on your machine. Local implementations support Windows/macOS/Linux, GPU acceleration where available, and CPU fallback. github.com

For a creator, it's a particularly good compromise between natural-sounding speech, relatively modest hardware requirements, and privacy. You download the model once, then your scripts can remain on the computer.

Kokoro offline implementation

2. Chatterbox — if expressiveness matters most.
I'd consider this if you're producing narration, character dialogue, or content where prosody/emotion matters more than minimal hardware requirements. It's more demanding than Kokoro, but the newer local TTS models generally offer substantially more expressive output than traditional lightweight systems.

3. Piper — best lightweight option.
Piper is a fast local neural TTS system with a large selection of languages and voices. It can run locally from downloaded ONNX voice models and is designed to be fast even on relatively modest hardware. github.com

Piper TTS

It's not quite as impressive as the newest expressive models, but if you're generating hours of narration and want something dependable, fast, and completely local, it's excellent.

What I'd choose

• YouTube/podcast narration: Kokoro
• Audiobook/character work: Chatterbox
• Lots of narration on an ordinary PC: Piper
• Voice cloning/experimental creator workflows: Fish Speech
• Highly sensitive scripts: any of the above running with the computer disconnected from the internet after the model is installed

One important distinction: "local" doesn't automatically mean the software never contacts the internet. Some front ends may check for updates or download models. For genuinely sensitive scripts, download the models beforehand and use a firewall/network block to verify that inference itself has no network access.

If you tell me your OS (Windows/Mac/Linux) and GPU (e.g. RTX 3060, Apple Silicon, or no GPU), I can narrow this down to the best 2–3 choices for your machine.

Yes. If privacy is the priority, I’d focus on software that runs the TTS model entirely on your computer rather than cloud services.

Best options for creators

My first choice: Voicebox. It's an open-source, local-first voice studio that runs on your machine and currently supports seven TTS engines, including Chatterbox, Qwen3-TTS, Kokoro, and LuxTTS. Its documentation explicitly says the processing happens on your hardware rather than in the cloud. docs.voicebox.sh

Voicebox documentation

For maximum voice quality: I'd pair Chatterbox with a good GPU. Current comparisons put it among the strongest local options for expressive speech and voice cloning, and its MIT license is particularly attractive if you're creating commercial content. offlinetts.com

For straightforward narration without cloning: Kokoro is an excellent choice. It's very small (82M parameters), fast, and can run locally without sending your scripts anywhere. offlinetts.com

One important caveat: don't automatically choose XTTS v2 just because it sounds excellent. Its CPML license has commercial-use restrictions, so it's a poor choice if you're producing monetized/commercial work unless your specific use is covered. www.promptquorum.com

If your scripts are genuinely sensitive

A local model gives you a much stronger privacy boundary: once the model files are downloaded, you can disconnect the machine from the internet and synthesize locally. I'd also disable automatic telemetry where applicable and keep your scripts/audio on an encrypted drive.

If you tell me Windows/Mac/Linux + your GPU (e.g. RTX 4060, Apple Silicon, etc.) + whether you need voice cloning, I can narrow this down to the best 2–3 choices and the easiest setup.

If your priority is high-quality voice synthesis without uploading scripts to a cloud service, there are several strong local options. The best choice depends on whether you need voice cloning or simply excellent narration with preset voices.

My picks

1. F5-TTS — best if you want a cloned/custom voice

F5-TTS is one of the local models I'd investigate first for a creator workflow. It can synthesize speech from a voice reference while keeping the actual script on your machine. It's particularly attractive if you have an NVIDIA GPU and want high-quality narration rather than a simple accessibility TTS engine.

2. Kokoro — best hassle-free private narration

Kokoro is excellent if you don't need to clone your own voice. Local implementations can run entirely offline after the models and voices have been downloaded; one implementation explicitly supports an offline mode with no subsequent internet connection required. github.com

It's probably the option I'd choose for YouTube/video narration where privacy matters but a particular cloned voice isn't essential.

3. XTTS-v2 — very good cloning, but watch the license

XTTS-v2 supports voice cloning from a short reference recording, multilingual synthesis, and 24-kHz output. github.com It can be run locally, so your scripts don't need to leave your computer.

However, there's an important catch for professional creators: XTTS-v2 is under the Coqui Public Model License (CPML), and current documentation/discussions indicate it isn't licensed for commercial use. github.com So I wouldn't automatically recommend it for monetized YouTube, commercial voiceover work, client projects, etc.

4. Piper — best for modest hardware

Piper is much less demanding and extremely practical for fully local TTS. The tradeoff is that its voices generally aren't as expressive or cinematic as the newer neural models.

If privacy is your main concern

The important distinction is local inference vs. "private mode" on a cloud service.

With a genuinely local setup:

script → your computer → model → WAV file

Your script, generated audio, and voice reference can all remain on your machine. Once the models are downloaded, you can even disconnect the computer from the internet. Kokoro's local implementations, for example, document exactly this workflow. github.com

I'd therefore rank them:

Best overall creator setup: F5-TTS
Best easy/private narration: Kokoro
Best multilingual cloning: XTTS-v2 (subject to its licensing restrictions)
Best low-resource option: Piper

One other consideration: if you're cloning your own voice, that's straightforward. If you're cloning someone else's voice, make sure you have their permission; the XTTS documentation itself cautions against using voice cloning to impersonate people without consent. github.comgithub.com

If you tell me Windows/macOS/Linux + your GPU (e.g. RTX 4070, Apple Silicon, or CPU-only), I can narrow this down to the best 1–2 options and explain how to set up a completely offline creator workflow.

If privacy is the priority, there are several strong offline/local options. The best choice depends on whether you want a polished creator app or are comfortable installing AI models yourself.

My top picks

1. LocalText2Voice — best overall for creators

LocalText2Voice on GitHub

This is probably the closest match to your use case. It's a desktop application for long-form narration and can run fully locally/offline, with engines including Kokoro, Piper, Chatterbox, Qwen3 TTS and OmniVoice. It also handles things like podcast-style mixing, normalization and exporting, rather than being merely a developer TTS engine. github.com

2. Kokoro — best balance of quality, speed and simplicity

Kokoro is an 82M-parameter local TTS model that can run without sending your script to a cloud service. A local implementation reports roughly 300 MB of model weights and CPU fallback, making it unusually accessible compared with larger voice models. github.com

Kokoro project

For YouTube narration, explainer videos, audiobooks and similar work, I'd try Kokoro first if you don't need voice cloning.

3. F5-TTS / Fish Speech — if you need voice cloning

These are more interesting when you want a particular voice rather than simply selecting a stock narrator. Local packages such as LocalSoundsAPI bundle XTTS v2, Fish Speech and Kokoro into an entirely local workflow. github.com

LocalSoundsAPI

4. Piper — best lightweight option

Piper TTS

Piper is extremely fast and designed specifically for local neural TTS. It has a large selection of languages/voices and can run on relatively modest hardware. The tradeoff is that its voices generally sound less sophisticated and expressive than the newer models. github.com

One important privacy distinction

"Offline" should mean the actual inference happens on your machine, not merely that an app has an offline mode. For genuinely sensitive scripts, I'd download the models once, disconnect the machine from the internet, and verify that the application still generates speech.

Also check the license of the particular voice/model if you're monetizing videos, audiobooks or other commercial work—Piper, for example, explicitly notes that individual voice models can have different licenses. github.com

If I were choosing for a creator with confidential scripts:
LocalText2Voice + Kokoro for straightforward narration; LocalText2Voice + F5-TTS/Chatterbox if voice cloning is important.

Yes. If privacy means the script and voice audio should never leave your computer, there are several good local options now.

My top picks

1. Voicebox — probably the easiest choice for creators.
It's a local-first voice studio that combines several engines, including Qwen3-TTS, Chatterbox, LuxTTS and Kokoro. It can clone voices, generate multilingual speech, apply effects, and keep the models and voice data on your machine. github.comgithub.com

Voicebox on GitHub

2. Chatterbox — my pick if natural/expressive delivery matters most.
The local version supports zero-shot voice cloning from a short reference recording, with Turbo, standard and multilingual models. Chatterbox Turbo also supports expressive tags such as [laugh] and [sigh]. github.com

Chatterbox on GitHub

3. Qwen3-TTS — particularly interesting for multilingual creators.
It can be run locally and is available through creator-friendly wrappers such as Voicebox. It's a strong option if you're producing content in multiple languages or want more control over delivery. github.comgithub.com

4. Kokoro — excellent if you don't need voice cloning.
It's dramatically smaller and easier to run than the larger models, making it attractive for laptops and CPU-only systems. Local implementations report an 82M model and fast CPU inference. github.comgithub.com

One important licensing warning

Don't assume "local" = "commercially unrestricted." For example, XTTS-v2 is quite capable and runs locally, but its current Coqui Public Model License has commercial-use restrictions; F5-TTS also has licensing limitations. github.com

For a YouTuber, podcaster, audiobook creator, or commercial video producer, I'd therefore start with Voicebox + Chatterbox or Qwen3-TTS, and verify the specific model's license before monetizing the output.

If your priority is maximum privacy + near-ElevenLabs quality, I'd rank them:
Chatterbox/Qwen3-TTS → Voicebox as the interface → Kokoro if you need something lightweight.

And "offline" should mean more than just the synthesis happens locally: download the model weights first, then disconnect the machine/network and verify that your workflow still works. That gives you much stronger assurance that private scripts aren't being transmitted. offlinetts.com

If privacy is the priority, I'd focus on software/models that run entirely on your own computer rather than cloud TTS APIs. The strongest options right now are:

🥇 Chatterbox — my first choice

Chatterbox by Resemble AI is particularly interesting for creators who don't want scripts leaving their machine.

It can run locally, supports zero-shot voice cloning from a short reference recording, and has controls for expressive/emotional delivery. The model is MIT licensed, and the company explicitly supports self-hosted/on-premise and air-gapped deployment. www.resemble.ai

The important privacy distinction is: if you install and run the model locally, your script can remain on your computer. You don't need to send the text to Resemble's cloud service. The model is available through its local Python package/Hugging Face distribution. huggingface.co

There's also Chatterbox Turbo, a smaller 350M-parameter version designed for very fast inference, with reported ~75 ms latency on a GPU. www.resemble.ai

🥈 F5-TTS — excellent alternative

F5-TTS on GitHub is another strong local option. It supports voice cloning from reference audio and provides both a local GUI and command-line inference. It can also run through Docker, making it practical to isolate the whole system from the internet. github.com

It's particularly attractive if you're comfortable with technical software and want maximum control over where the data goes.

🥉 Kokoro — if you don't need cloning

Kokoro is worth considering if you mainly want natural narration using preset voices rather than reproducing your own voice. It's considerably lighter than many cloning models and is well suited to running locally.

What I'd choose

• YouTube/video narration + expressive delivery: Chatterbox
• Cloning your own voice: Chatterbox or F5-TTS
• Lots of narration on modest hardware: Kokoro
• Technical user who wants maximum control: F5-TTS
• Extremely sensitive scripts: download the model once, then disconnect the machine from the internet while generating

One caveat: "offline" only guarantees privacy if the particular application actually performs inference locally. A desktop app can use a local model while still sending telemetry or other information online. For highly private material, I'd favor the open-source model + local interface approach and optionally block the application from network access.

If you tell me your computer (Windows/Mac/Linux + GPU, if any) and whether you want your own voice cloned or a stock narrator, I can narrow this down to the best 1–2 options for your setup.

Yes. If privacy is the priority, I’d focus on models that run entirely on your own computer after the initial model download. That means your scripts never need to leave the machine.

Best offline options

My first choice: Kokoro. Kokoro-82M is an open-weight model with a relatively small footprint, and implementations can run completely offline once the model and voices are downloaded. It supports multiple languages/voices and is particularly attractive when you want good narration without needing a huge GPU. github.com

Kokoro TTS project

For voice cloning: Chatterbox. Chatterbox is open source and its current Multilingual V3 model is a 0.5B-parameter model aimed at natural, expressive speech and speaker similarity. Its Turbo variant can clone from a reference audio clip and supports expressive tags such as laughter and coughing. github.com

Chatterbox TTS on GitHub

For simplicity and speed: Piper. Piper is explicitly designed as a fast local neural TTS system. It runs models locally, has a large selection of languages and voices, and can output WAV directly—making it good for batch-producing narration without sending scripts to a cloud service. github.com

Piper TTS

What I'd choose for a creator

• YouTube/podcast narration: Kokoro
• Character voices / voice cloning: Chatterbox
• Lots of scripts + fast batch generation: Piper
• You have a powerful GPU and like tinkering: Chatterbox or StyleTTS 2
• Absolute privacy: download the models once, then disconnect the machine from the internet while synthesizing.

One important distinction: "open source" doesn't automatically mean every voice/model can be commercially used. Check the specific model and voice license before using generated audio commercially; Piper, for example, explicitly notes that individual voices can have different licensing restrictions. github.com

If you tell me your OS (Windows/macOS/Linux), GPU (e.g. NVIDIA RTX 4070 / Apple Silicon), and whether you need voice cloning, I can narrow this down to the best 1–2 setups for you.

If privacy is the priority, I’d focus on tools where the model runs entirely on your own computer rather than merely services that promise not to retain uploads.

Best options

My first choice for a creator: Coqui XTTS-v2. It can clone a voice from a short reference recording, supports multilingual generation, and can be run locally through the open-source TTS toolkit. github.com

Coqui TTS / XTTS-v2 on GitHub

OpenVoice V2 is another strong option, particularly if you want voice/style control and multilingual output. Its V1/V2 code is MIT-licensed and the project documents local installation. github.comgithub.com

OpenVoice V2 on GitHub

What I'd use for private creator work

• Narration / YouTube / podcasts + your own voice: XTTS-v2
• Voice cloning + experimenting with accents/styles: OpenVoice V2
• Don't need cloning, just excellent synthetic narration: Kokoro or Piper
• Very sensitive scripts: download the model once, then disconnect the machine from the internet during synthesis. That's stronger privacy than relying on a provider's cloud-storage policy.

One caveat: "offline" applies to the inference process. You may still need internet initially to download the software/model, and some surrounding GUI applications can make network requests. It's worth checking the actual network behavior rather than assuming an app is private because it advertises itself as offline. offlinetts.com

If you tell me your OS (Windows/macOS/Linux) and GPU, I can narrow this down to the best-quality option that will actually run well on your machine.

If privacy is the priority, I’d focus on software that runs the TTS model entirely on your computer after the initial model download.

My top pick: Kokoro

Kokoro is probably the sweet spot for a creator who wants high-quality narration without sending scripts to a cloud service. The 82M-parameter model can run locally, and local implementations support multiple voices and languages. After downloading the model/voices, you can configure it to operate completely offline. github.com

For a more polished creator workflow, Kokoro Studio is a particularly interesting option: its current local GUI supports GPU acceleration, audiobook mode, voice mixing, project saving, and EPUB input while keeping synthesis offline. github.com

Kokoro Studio on GitHub

Best lightweight option: Piper

Piper is excellent if you value speed, low hardware requirements, and straightforward offline operation over maximum expressiveness. It runs locally, supports a large collection of languages/voices, and can generate WAV files without sending your script anywhere. github.com

Piper TTS on GitHub

If you want voice cloning

That's where I'd look at XTTS/Coqui-derived local setups or other locally hosted voice-cloning models. They can be considerably more expressive and allow custom voices, but they're more demanding to configure and you need to pay particular attention to the voice/model license and consent for any voice you clone.

Privacy caveat

"Offline" should mean more than the company says it is private. For genuinely sensitive scripts, I'd:

1. Download the model and voices.
2. Disconnect the machine from the internet or block the application's network access.
3. Verify it can synthesize with networking disabled.
4. Keep scripts and generated audio on an encrypted local drive.
5. Check the specific voice/model license, especially for commercial work.

Bottom line: For most creators, I'd try Kokoro first, particularly a local GUI such as Kokoro Studio. Choose Piper if you want something faster/lighter and don't need as much expressive narration. github.comgithub.com

If you tell me whether you're on Windows, macOS, or Linux and whether you have an NVIDIA/Apple GPU, I can narrow this to the best 2–3 options for your hardware.

If privacy is the priority, I’d focus on software/models that perform inference entirely on your own computer after the initial model download.

Best local options

My first choice: Kokoro. It's one of the strongest options if you want natural-sounding narration without sending scripts to a cloud service. Local implementations can run completely offline once the model and voices have been downloaded. github.com

For voice cloning, look at Chatterbox. Current 2026 comparisons put it among the strongest open/local alternatives to cloud TTS, particularly when cloning and expressive speech matter. merchantprotocol.com

XTTS-v2 is another very capable choice, especially if you need multiple languages and voice cloning; it supports 16 languages and can run inference locally. github.com Important caveat for commercial creators: XTTS-v2's model is under the Coqui Public Model License, which does not permit commercial use, and there currently isn't a way to purchase a commercial license. github.com

Piper is the practical lightweight option. It is a fast local neural TTS system and can generate WAV files entirely from locally stored models. github.com

If your scripts are genuinely confidential

Don't just look for a product advertised as "privacy-focused." Verify that:

• the actual synthesis happens locally;
• scripts aren't uploaded for processing;
• voice samples aren't uploaded to a server;
• telemetry can be disabled;
• the model files can be downloaded and then used with the network disconnected.

For maximum assurance, disconnect the computer from the internet after downloading the models and test the software. "Offline" implementations can still perform downloads or other network requests during initial setup. offlinetts.com

Bottom line:

• 🎙️ Best overall local narration: Kokoro
• 🎭 Best if you need voice cloning: Chatterbox
• 🌍 Best multilingual cloning: XTTS-v2, but check its non-commercial license
• 💻 Best lightweight/low-resource: Piper

If you tell me whether you're on Windows, Mac, or Linux and roughly what GPU/RAM you have, I can narrow this down to the best 2–3 options that can produce near-ElevenLabs-quality voices locally.

Yes. If privacy is the priority, there are now several genuinely local options where your script can stay on your computer.

Kokoro is particularly attractive for private narration because it's small enough to run locally and doesn't require sending scripts to a server. socialstream.ninja

For voice cloning, I'd look first at Chatterbox or Qwen3-TTS. Current local-TTS comparisons rate both highly for natural cloning, while Chatterbox is MIT-licensed and Qwen3-TTS has Apache 2.0 licensing according to recent comparisons. localaimaster.com

My picks by creator type

• “I just want ElevenLabs-like narration without uploading my scripts” → OpenVox or Kokoro.
• “I want to clone my own voice locally” → Chatterbox.
• “I want the newest local voice-cloning technology and don't mind technical setup” → Qwen3-TTS.
• “I have a modest/CPU-only computer” → Piper or Kokoro.
• “I want a production workflow rather than command-line tools” → OpenVox. It specifically supports scripts, PDFs/EPUBs, long-form audio, voice cloning, and local processing. openvoxai.com

One important caveat: “offline” doesn't automatically mean every downloaded voice/model is commercially licensed. If you're monetizing YouTube videos, audiobooks, games, etc., check the specific model/voice license. For example, F5-TTS and XTTS v2 have licensing restrictions that can matter commercially. www.promptquorum.com

If you tell me Mac/Windows/Linux + your GPU (if any), I can narrow this down to the 2–3 best options that will actually run well on your machine.

If privacy is the priority, I’d look at local/offline TTS rather than creator platforms that process scripts in the cloud. As of 2026, these are the strongest options:

My picks

1. Kokoro — best overall for private narration

Kokoro-82M is unusually small while producing very natural speech. It can run locally, including on CPU, and its Apache 2.0 licensing is creator-friendly. The major limitation is that you choose from its available voices rather than cloning your own. localaimaster.com

2. Chatterbox — best if you want your own voice

Chatterbox is a strong choice when you want high-quality voice cloning while keeping the script and reference audio on your machine. Its model is MIT-licensed, although you should still check the particular model/version and your intended use. localaimaster.com

3. F5-TTS — excellent local cloning, but watch the license

Technically, F5-TTS is one of the most impressive local cloning options: give it a short voice sample and it can synthesize new speech locally. However, current sources flag the F5-TTS weights as non-commercial (CC-BY-NC), so I'd avoid it for monetized creator work unless the applicable license for your specific version permits it. offlinetts.comlocalaimaster.com

4. Piper — best if your computer isn't powerful

Piper is exceptionally lightweight and CPU-friendly. Quality isn't quite at Kokoro/Chatterbox level, but it's excellent when you want something that can run entirely offline without a powerful GPU. Be careful about the license of the particular Piper fork and individual voice you use. offlinetts.comlocalaimaster.com

An easy option for nontechnical creators

There are also desktop/browser interfaces built around these models. OfflineTTS, for example, runs English Kokoro synthesis locally in the browser after downloading the model; its documentation says the English TTS processing happens on-device. offlinetts.comofflinetts.com

One important privacy caveat: “offline” should mean you can disconnect from the internet and still synthesize new audio. Some tools download models locally but send text to a server for preprocessing, particularly for certain languages. offlinetts.comofflinetts.com

Bottom line:

• Private YouTube/podcast narration: Kokoro
• Private cloning of your own voice: Chatterbox
• Maximum quality experimentation: F5-TTS, if non-commercial licensing is acceptable
• Older/low-power PC: Piper

If you tell me your OS (Windows/macOS/Linux), GPU (if any), and whether you need voice cloning, I can narrow this down to the best 1–2 options and explain how to run them completely offline.

Yes. If the priority is high-quality voice synthesis while keeping scripts entirely on your machine, I’d look at local/open-weight TTS rather than cloud services.

My top picks

1. F5-TTS — best overall for serious creators
F5-TTS official GitHub

F5-TTS can run entirely locally, supports voice prompting/cloning from reference audio, and has both a command-line interface and local web UI. Its documentation explicitly supports offline PyTorch inference and even Apple Silicon. github.com

2. Chatterbox — best for expressive voices
Chatterbox official GitHub

Chatterbox is open-source and runs locally. Its newer models support voice cloning and expressive/paralinguistic effects such as laughs and chuckles; the multilingual V3 model is aimed at natural conversational speech. github.com

3. Kokoro — best lightweight option
Kokoro official GitHub

Kokoro is only 82M parameters, has Apache-licensed weights, and is designed to deliver surprisingly high quality while being much faster and cheaper computationally than larger models. It can run locally on Windows, Linux, and Apple Silicon. github.com

If privacy is the deciding factor

A local model gives you a much stronger privacy boundary: your script can remain on your computer and the generated audio can be produced without sending the text to a TTS API. You still need to be careful about the installation itself—download the software/models from trusted sources, and ideally run the application with network access disabled if you're dealing with particularly sensitive material.

For a creator, I'd choose F5-TTS if you have a reasonably powerful GPU, Chatterbox if expressive/character performance matters most, and Kokoro if you want something considerably lighter and easier to run.

If you tell me whether you're on Windows, macOS, or Linux and what GPU/CPU you have, I can narrow this down to the best local setup for your machine.

If privacy is the priority, I’d favor local/self-hosted TTS models where your script never needs to leave your computer.

My top picks

1. Chatterbox — best overall for creators

Chatterbox on GitHub

Chatterbox is an open-source TTS model from Resemble AI, licensed MIT. It supports zero-shot voice cloning from a short reference recording and includes controls for expressive delivery. The model can be downloaded and run locally, so your actual scripts can remain on your machine. github.com

For YouTube, podcasts, audiobooks, and video narration, this would be my first thing to try if you have a reasonably capable GPU.

2. F5-TTS — excellent alternative

F5-TTS on GitHub

F5-TTS is another strong open-source voice-cloning system. Its official implementation supports local installation, CLI inference, Docker, and offline inference. github.com

It's particularly interesting if you don't mind a somewhat more technical setup.

3. Kokoro — best lightweight option

Kokoro TTS Local on GitHub

Kokoro-82M is dramatically smaller than many modern voice-cloning models. A local implementation supports 54 voices across eight languages and can operate completely offline after the initial model download. github.com

I'd choose Kokoro if privacy + low hardware requirements + fast narration matter more than cloning your exact voice.

4. Voicebox — best if you want an actual creator-oriented app

Voicebox documentation

Voicebox is a local-first voice studio rather than just a model. It bundles multiple TTS engines—including Chatterbox and Kokoro—and is designed to keep processing on your hardware. docs.voicebox.sh

What I'd choose

• Maximum voice quality/cloning: Chatterbox
• High quality + technical flexibility: F5-TTS
• Lower-end computer / straightforward narration: Kokoro
• Want a GUI/workflow rather than tinkering with Python: Voicebox

One important privacy distinction: "has an offline mode" isn't quite the same as "never connects to the internet." For genuinely confidential scripts, download the models first, then disconnect the machine/network or enforce offline mode and verify that the application isn't making outbound requests. For example, Kokoro-TTS-Local explicitly documents an offline mode after its models and voices have been downloaded. github.com

If you tell me Windows/macOS/Linux and your GPU (e.g. RTX 4070, Apple Silicon, no GPU), I can narrow this down to the best 1–2 options for your hardware.