Create natural-sounding voice-overs with open AI models that run in your web browser. Choose VoxCPM2 for 30 languages, including Khmer, Thai, Vietnamese, Chinese and English, and describe the voice you want ("a calm middle-aged man, deep voice"), or pick Kokoro, a small and fast English model with nine ready-made American and British voices. Download the result as a WAV or MP3 file for videos, presentations, lessons or podcasts.
Everything is generated on your own device: the text never leaves your computer and there is no account, character limit per month or watermark. The models download once and are cached in your browser. For the simple built-in voices of your system, use the classic text to speech reader; to speak in your own voice, try AI voice cloning.
How to use the AI Text to Speech tool
- Type or paste your text (up to 5,000 characters). It is split into sentences automatically.
- Choose a model: VoxCPM2 (multilingual, needs a GPU) or Kokoro (English, works on any computer).
- For VoxCPM2, optionally describe the voice; for Kokoro, pick a voice and speed.
- Press "Generate speech". The model downloads the first time and is cached for next time.
- Listen as each sentence is ready, then download WAV or MP3.
Why use ToolOrca AI Text to Speech
- Two open models: VoxCPM2 (30 languages, 48 kHz) and Kokoro-82M (English, 24 kHz)
- Khmer, Thai, Vietnamese, Chinese, Japanese, Korean and more with VoxCPM2
- Voice design from a short description, or keep one voice for the whole text
- Runs on your GPU with WebGPU; Kokoro also runs on the CPU
- Live playback, Stop button, WAV and MP3 download
- No upload, no sign-up, no watermark
Hardware requirements
VoxCPM2 needs WebGPU (recent Chrome or Edge on a computer, or Safari 26 on a Mac) and about 2–3 GB of free GPU memory; the first run downloads about 1.6 GB (plus 190 MB for the voice lock). An Apple Silicon Mac or a PC with a graphics card produces roughly real-time speech; without WebGPU it is far too slow.
Kokoro is only 82 million parameters: it downloads about 90 MB for the CPU or 330 MB for the GPU and runs on almost any computer, and on many phones, even without WebGPU, a few times faster than real time on a GPU.
Which model should I use?
Use Kokoro for English when you want speed and a consistent, polished voice. Use VoxCPM2 for every other language, especially Khmer and other Southeast Asian languages, or when you want to design a voice from a description. VoxCPM2 picks a new voice for each run unless you describe one; leave "Keep the same voice for every sentence" on so long texts stay in one voice.
Tips: write numbers and abbreviations as words for the most reliable reading, keep sentences moderately short, and generate again if a sentence sounds odd, as every run is a little different.
Frequently asked questions
Does it support Khmer text to speech?
Yes. Choose the VoxCPM2 model and type Khmer text. It also speaks Thai, Lao, Burmese, Vietnamese, Chinese, Japanese, Korean, English and about 20 more languages.
Is my text sent to a server?
No. The AI model runs in your browser on your own device. Only the model files are downloaded from Hugging Face, once, and stored in your browser.
Can I use the audio in my videos?
Both models are released under the Apache 2.0 licence, and the audio you generate is yours to use, including in commercial videos. Do not use it to impersonate real people.
Why is VoxCPM2 not available on my phone?
It needs WebGPU and several gigabytes of graphics memory, which most phones do not have yet. Use the Kokoro model on phones, or a computer for VoxCPM2.
How is this different from the normal Text to Speech tool?
The Text to Speech tool reads text with the voices built into your operating system and cannot save a file. This page uses neural AI models for more natural voices and lets you download WAV or MP3.