If you are looking for a lightweight AI voice model that actually sounds good, I found six that genuinely impressed me. Small, open source, and free to run locally. Some need almost no GPU at all. One is literally 25MB.
For their size they sound closer to paid platforms like ElevenLabs than anyone would expect.
Table of contents
1. KittenTTS
I almost skipped this one because 15 million parameters sounds like too little for a TTS model. Then I listened to it.
It is not perfect but for something that fits in 25MB and runs on a CPU with absolutely no GPU it is genuinely surprising. Eight built in voices, real time inference, and it works on literally any device that runs Python.
There are three variants worth knowing. The nano at 15M parameters and 25MB is the one that breaks your expectations. The micro at 40M sits in the middle. The mini at 80M is the most capable of the three and still smaller than most apps on your phone.
Features of KittenTTS
- 15M parameters, 25MB in the smallest variant
- Runs on CPU with zero GPU required
- Eight built in voices including Bella, Jasper, Luna, Bruno and more
- Real time inference optimized
- Apache 2.0 licensed
VRAM requirements: Zero. Runs entirely on CPU.
Best for
- Developers who need the lightest possible local TTS
- Edge deployment, low resource environments, any device
2. Kokoro 82M
82 million parameters & Trained for roughly $1000 on a few hundred hours of audio. And it consistently ranks at the top of open source TTS arenas.
That cost detail is not just trivia. It tells you something about how efficient this architecture is. Most models at this quality level cost hundreds of thousands to train. Kokoro did it for less than a used car.
The voice quality is where it actually surprises you. It does not sound like a lightweight model. Naturalness, pacing, prosody, it handles all of it better than models three or four times its size. 54 voices across 8 languages give you enough variety for most real world use cases.
Apache 2.0 licensed, already deployed in many products, and available via API for under $1 per million characters if you do not want to run it locally.
Features of Kokoro 82M
- 82M parameters, significantly faster than comparable quality models
- 54 voices across 8 languages
- Apache 2.0 licensed
- Under $1 per million characters via API
- Runs on modest hardware
VRAM requirements: Low. Runs comfortably on modest GPUs and CPU inference is possible.
Best for
- Production deployments that need quality without heavy compute
- Developers who want commercial friendly licensing with no compromises on voice quality
3. LuxTTS
Most TTS models output at 24khz. LuxTTS outputs at 48khz. That difference is immediately noticeable, the audio just sounds cleaner and more detailed than what you are used to hearing from local models.
The speed is the other thing that gets me. 150x realtime on a single GPU. That means a one minute audio clip generates in under a second. It also runs faster than realtime on CPU which puts it in a different category from most voice cloning models.
Voice cloning needs just a 3 second reference clip. The community has already built Gradio apps, a clean UI called OptiSpeech, and even ComfyUI nodes around it. For a model with 1.8K stars that is a healthy ecosystem.
Float16 inference is still coming which should make it nearly 2x faster. Already fast enough that it barely matters.
Features of LuxTTS
- 48khz audio output vs the standard 24khz
- 150x realtime speed on GPU, faster than realtime on CPU
- Voice cloning from a 3 second reference clip
- Runs on GPU, CPU, and Apple Silicon MPS
- Apache 2.0 licensed
VRAM requirements Fits within 1GB VRAM. Works on any local GPU.
Best for
- Creators who want the cleanest local audio quality
- Voice cloning without heavy hardware requirements
Related: Open-Source AI Text-to-Speech Generators You Can Run Locally for Natural, Human-Like Voice
4. CosyVoice 2
If you are building anything that needs real time voice like a voice bot, a streaming assistant, anything that has to respond fast, CosyVoice2 is the one to look at.
150ms latency. It supports both text input streaming and audio output streaming simultaneously which is what makes that latency possible. Most models process the full input before generating any audio. CosyVoice2 starts generating while text is still coming in.
The benchmark numbers back it up too. At 0.5B parameters it holds its own against models three times its size on speaker similarity and content consistency. The RL version of Fun-CosyVoice3 which builds on this architecture actually beats most 1.5B models on the hard test set.
Nine languages supported including Chinese dialects, emotion control, speed and volume instructions, and zero shot voice cloning all in a 0.5B package.
Features of CosyVoice2 0.5B
- 150ms ultra low latency streaming
- Bi-streaming — text in and audio out simultaneously
- Zero shot voice cloning
- 9 languages plus 18 Chinese dialects
- Emotion, speed and volume control via instructions
VRAM requirements Consumer GPU recommended for smooth streaming inference.
Best for
- Voice bots and real time conversational AI
- Streaming applications where latency matters
5. MeloTTS
Most lightweight TTS models pick one language and do it well. MeloTTS covers English, Spanish, French, Chinese, Japanese and Korean in a single small model and the English support alone comes in four accents: American, British, Indian and Australian.
That accent variety is the part I did not expect. Most models treat English as one thing. MeloTTS treats it as four distinct speakers which matters if your product has a global audience.
MIT licensed, built by researchers from Tsinghua University and MIT. Fast enough for CPU real time inference which means no GPU required at all. Just install, pick your language and accent, and generate.
The Chinese speaker also handles mixed Chinese and English naturally which is genuinely useful for anyone building for that audience.
Features of MeloTTS
- 6 languages including 4 English accents
- CPU real time inference, no GPU needed
- Mixed Chinese and English support
- MIT licensed, free for commercial use
- Simple API, minimal setup
VRAM requirements Zero. Runs entirely on CPU in real time.
Best for
- Multilingual products that need one lightweight model for multiple languages
- Developers who want accent variety without running separate models
So which one is right for you?
- KittenTTS: The one to grab if size and hardware constraints are the priority.
- Kokoro 82M: Best quality to size ratio on this list. Nothing at this weight sounds this good right now.
- LuxTTS: Voice cloning with the cleanest audio output. 48khz and 150x realtime speed in under 1GB VRAM is genuinely impressive.
- CosyVoice2 0.5B: Built for real time. If latency matters for what you are building this is the one.
- MeloTTS: Six languages, four English accents, runs on CPU. The obvious pick for multilingual products.
All five are free, open source, and need no subscription or API key. That is the real win here.
