Open-source text-to-speech has become good enough that you no longer need a paid voice service for many jobs: narrations, prototypes, accessibility, quick voice-overs. The catch is usually the setup: Python environments, CUDA, model downloads. Below are six recent open models you can try straight from the browser, with what each one is best at.
| Model | Languages | Voice cloning | Size | License |
|---|---|---|---|---|
| Qwen3-TTS 1.7B | 10 (incl. Italian, Russian, Portuguese) | Yes, plus voice design from a description | 1.7B | Apache-2.0 |
| Chatterbox Multilingual | 23 | Yes, plus an expressiveness setting | 0.5B | MIT |
| Ming 3.0 TTS | 7 (EN, ZH, JA, KO, FR, DE, ES) | Yes, plus "describe a voice" | 1.5B | MIT |
| Audio8 TTS Realistic | 23 | Yes, zero-shot from a short clip | 0.6B | Apache-2.0 |
| Phonon 31M | English | No, 10 voices incl. characters | 31M (64 MB) | Apache-2.0 |
| Paradee 1.2 | English | No, 4 built-in voices | 8M (8.5 MB) | Apache-2.0 |
Qwen3-TTS: the most complete
Alibaba's Qwen team released three 1.7B checkpoints, one per job: nine built-in voices that take a style instruction ("very happy", "whispering"), a VoiceDesign model that creates a voice from a written description, and a Base model for cloning. Every voice speaks all 10 languages, Italian included. It is the slowest here: on our GPU a short sentence takes about 30 seconds.
Qwen3-TTS 1.7BAlibaba Qwen · Apache-2.0Alibaba's open text-to-speech in 10 languages: nine built-in voices with emotion control, voices designed from a description, and cloning.Try it online →Chatterbox Multilingual: fast cloning in 23 languages
From Resemble AI, MIT-licensed. It clones a voice from a short clip and has a simple expressiveness setting, from calm to dramatic. It is quick: in our tests a short Italian sentence was ready in under 3 seconds. Every file carries Resemble's inaudible Perth watermark, as in the official release.
Chatterbox MultilingualResemble AI · MITResemble AI's open voice model in 23 languages: clone a voice from a short clip and turn the expressiveness up or down.Try it online →Ming 3.0 TTS: the most expressive
From inclusionAI (Ant Group). You can pick a built-in voice, describe a voice in words ("a calm, deep male voice") or clone a short reference clip. Tags such as [Laughter] or [Sigh] in the text add expression. It speaks English, Chinese, Japanese, Korean, French, German and Spanish; the built-in voices speak English.
Audio8 TTS Realistic 0.6B: the most languages
From Edge0. It covers 23 languages and clones a voice from a short clip (up to 15 seconds, plus its transcript). It is the slowest of the three: on our GPU it generates at roughly 0.4× real time, so a 10-second sentence takes around 25 seconds.
Audio8 TTS Realistic 0.6BEdge0 · Apache-2.0Realistic text-to-speech in 23 languages, with zero-shot voice cloning from a short clip.Try it online →Phonon 31M: the newest, now in Ollama
From Gradium, and just added to Ollama (on Apple Silicon Macs only, for now). 31M parameters in a 64 MB file, running on a native runtime without PyTorch, so we run it on a CPU too. It reads numbers, prices and dates naturally, and its voices include characters such as an old pirate, a wizard and a dragon. More in our Phonon guide.
Phonon 31MGradium · Apache-2.0 (31M English model and voices)A 31M-parameter English text-to-speech model with ten voices, from narrators to a pirate and a dragon: runs on a plain CPU.Try it online →Paradee 1.2: the smallest
An 8-million-parameter English model distilled from Kokoro, only 8.5 MB. It is so small that we run it on a plain CPU, with no GPU at all, and it answers in a second or two. Four voices: American and British, female and male, with adjustable speed. Perfect when you need fast, lightweight English speech.
Paradee 1.2 (8M)Sahil Mahendrakar · Apache-2.0A tiny 8M-parameter English text-to-speech model distilled from Kokoro: 8.5 MB, faster than real time on one CPU thread.Try it online →Which one should I use?
- A voice designed from a description, in 10 languages: Qwen3-TTS.
- Fast cloning in 23 languages, with more or less emotion: Chatterbox.
- Expressive narration with tags like [Laughter]: Ming 3.0 TTS.
- A language other than the main seven, or cloning your own voice: Audio8 TTS.
- English narration or character voices on a CPU: Phonon 31M.
- The smallest possible English model, to embed in an app: Paradee.
All six allow commercial use of the audio under their licenses (MIT or Apache-2.0). Only clone voices you have the right to use: LiveWeights asks you to confirm consent before uploading a reference clip.
And speech-to-text?
For the opposite direction, Underdog Bark 3.0 turns English audio into a timed transcript: record from the microphone or upload a file.