Open-source text-to-speech models can narrate articles, create video voiceovers, give game characters distinct voices, and make AI assistants speak. Many run locally and can generate speech in several languages, clone a recorded speaker, or create a new voice from a text description.
The seven models and model families featured here are useful for those projects, including everyday narration, custom voices, long-form recordings, podcasts, and live conversations. You can listen to official audio examples and follow the model links to try them on your own device.
Last Updated Oct 12, 2026
Quick comparison
| Model | Main use | Local hardware |
|---|---|---|
| Kokoro | Preset-voice narration | CPU or GPU |
| Qwen3-TTS | Voice cloning and voice design | GPU preferred |
| Chatterbox | Multilingual cloning and expressive speech | CPU (Nano), GPU (other variants) |
| VoxCPM2 | Custom voices and expressive cloning | GPU, or CPU via GGUF |
| CosyVoice 3 | Multilingual speech and pronunciation control | GPU preferred |
| MOSS-TTS | Long-form audio, dialogue, voice agents | CPU (Nano), GPU (larger models) |
| dots.tts | Voice cloning and speech editing | GPU preferred |
1. Kokoro: lightweight narration with preset voices
Kokoro-82M reads text aloud using a collection of preset voices. Its compact 82-million-parameter model is well suited to article narration, offline readers, accessibility features, and video voiceovers that need a consistent narrator.
It runs on CPUs and GPUs and generates 24 kHz audio. The Python library lets you choose a voice, change speaking speed, and generate longer scripts in chunks. Available voices include American and British English, Spanish, French, Hindi, Italian, Japanese, Brazilian Portuguese, and Mandarin Chinese.
Kokoro works particularly well when one of the supplied voices suits the project. The standard model does not clone a speaker from a sample recording or design a voice from a text prompt. For a local reading app or a set of tutorials with a consistent narrator, the small model size makes it a convenient place to begin.
Audio sample: Preset English voice.
License: Apache-2.0
2. Qwen3-TTS: voice cloning, preset voices, and voice design
Qwen3-TTS can reproduce a voice from a recording, generate speech with supplied speakers, or create a new voice from a written description. The family supports English, Chinese, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian.
- Base: Clone a speaker from reference audio or fine-tune a voice for repeated use.
- CustomVoice: Generate speech with nine supplied voices. The 1.7B model also accepts style instructions.
- VoiceDesign: Describe the voice you want in text and generate it with the 1.7B checkpoint.
These options work well for character dialogue, multilingual voiceovers, and narrated content that needs a recognizable voice. A generated VoiceDesign sample can also serve as reference audio for subsequent Base generations.
Qwen provides 0.6B and 1.7B checkpoints for the relevant tasks, along with the qwen-tts Python package and streaming speech generation. Choose the checkpoint for the voice task you need, since VoiceDesign is only available in the 1.7B release.
Audio sample: English voice cloning.
License: Apache-2.0
3. Chatterbox: voice cloning for multiple languages and expressive English speech
Chatterbox is Resemble AI’s TTS family for generating speech in a reference speaker’s voice. It works for dubbed videos, character lines, and spoken responses in voice applications. Different versions give you multilingual speech, expressive English delivery, or a smaller model for local devices.
- Multilingual V3: Voice cloning and speech generation across 23 languages.
- Turbo: A 350M English model with vocal-event tags including
[laugh]and[cough]. - Nano: A 110M English model for CPU and on-device speech generation.
- Single Language Pack: Dedicated models for Chinese, Hindi, Spanish for Latin America and Spain, and Portuguese for Brazil and Portugal.
The original and multilingual models also have settings for expression and adherence to the reference voice. Those are useful for scenes that need more energy or a more restrained reading.
Audio sample: English speech generation.
License: MIT
4. VoxCPM2: design a voice and control how it speaks
VoxCPM2 generates a new voice from a text description and can clone a speaker from recorded audio. Voice instructions can describe delivery, pace, and emotion, which is useful for game characters, animation, and voiceovers that need more expressive control.
The 2B-parameter model supports 30 languages and produces 48 kHz audio. For cloning, you can supply a recording and guide the speech style with text. A continuation workflow also uses the reference recording’s transcript when recreating a speaker.
Python inference supports batch and streaming generation. A GGUF-based C++ implementation can run on CPUs and on compatible GPU backends, although performance depends on the model files and hardware. VoxCPM2 is worth exploring when creating a voice is as important as generating the lines it will speak.
Audio sample: English speech.
License: Apache-2.0
5. CosyVoice 3: multilingual speech with pronunciation control
CosyVoice 3 generates speech in nine languages and supports more than 18 Chinese dialects and accents. Its zero-shot voice cloning works across languages, making it useful for multilingual dubbing and narration in a consistent speaker’s voice.
A useful feature is pronunciation editing. You can supply Chinese Pinyin or English phonemes for names and terms that need a specific reading. Instruction controls also adjust delivery, including emotion, speaking rate, and volume. That helps when a script contains product names or unusual terminology.
The 0.5B Fun-CosyVoice3 checkpoint supports streamed audio generation and incremental text input, which also makes it relevant to conversational voice agents. The official Python setup is GPU-oriented and includes the speech and text-processing dependencies.
Audio sample: Expressive English speech.
License: Apache-2.0
6. MOSS-TTS: narration, podcasts, and real-time speech
MOSS-TTS is a family of speech models built for narrated content, speaker cloning, dialogue, and live voice applications. Its different releases make it possible to work on a full-length recording, produce a scripted conversation, or run a compact voice generator locally.
- MOSS-TTS-v1.5: Multilingual long-form narration and voice cloning, with controls for pauses.
- MOSS-TTSD: Multi-speaker speech for scripted conversations and podcast-style audio.
- MOSS-TTS-Realtime: Incremental speech generation for voice agents.
- MOSS-VoiceGenerator: Voices and expressive delivery created from descriptions.
- MOSS-TTS-Nano: A compact 100M model for multilingual cloning and streaming speech on a CPU.
For a podcast with multiple speakers, MOSS-TTSD lets you assign the dialogue to speakers and use their recorded voices. For a reading app, Nano provides a smaller local option. The larger MOSS-TTS models use more demanding GPU setups, including 8B and 4B long-form checkpoints.
Audio sample: MOSS-TTS Nano Output.
License: Apache-2.0
7. dots.tts: multilingual cloning and speech editing
dots.tts is a 2B-parameter TTS family for generating speech from a reference voice. It produces 48 kHz audio and has several checkpoints for cloning, local inference, and fine-tuning.
For high speaker similarity, the SOAR checkpoint is the main cloning option. MeanFlow checkpoints reduce generation steps for quicker synthesis and serving. The model also supports an audio-only speaker reference or a continuation workflow that uses both a recording and its transcript.
The related dots.tts.edit model can edit speech you have already recorded. You can replace spoken words, change timing or expression, and adjust pauses while retaining the rest of the recording. The editing checkpoint is useful when a voiceover script changes after recording.
The Python package includes inference and fine-tuning code, with GPU-oriented serving options. For streaming applications, the available backend can emit audio chunks from supplied text.
Audio sample: English cloned speech.
License: Apache-2.0
How to evaluate the audio
Listen to the demo voices first, then try a few lines from the script you plan to use. Include names, abbreviations, numbers, and sentences with different punctuation. For a cloned voice, compare several generated lines with the reference recording. For dialogue and long narration, listen to a full passage to judge pacing, pronunciation, and consistency between speakers or sections.
FAQs
Can open-source TTS models run entirely offline?
Yes. These models can generate audio locally after you download their weights and required dependencies.
Do I need a GPU for text-to-speech?
Not for every model. Kokoro and MOSS-TTS-Nano have CPU inference options, and Chatterbox Nano focuses on CPU deployment. Larger voice-cloning and voice-design checkpoints generally benefit from a compatible GPU.
How much audio do I need to clone a voice?
It varies by model. Qwen3-TTS Base supports cloning from a short reference clip of around three seconds. Other models use their own reference-audio workflows, sometimes with a transcript. A clear recording of the intended speaker is a good starting point.
Related resources
- Privacy-First Local Text to Speech App – OpenWebTTS
- Voicebox: Free Local Voice Cloning, TTS, Dictation Studio
- Free Multilingual TTS & Voice Clone That Runs on CPU – MOSS-TTS-Nano
- 7 Best Free AI Voice Cloning Tools
