AI agents · OpenClaw · self-hosting · automation

Quick Answer

Best Open-Source Text-to-Speech Models 2026: 7 Compared

Published:

The short answer

Use Kokoro for cheap, fast narration; Chatterbox or Qwen3-TTS for voice cloning; and VibeVoice-Realtime or Qwen3-TTS for real-time voice agents. Open-weight TTS models now cover most jobs that needed a paid API in 2025, but licences vary more than quality does. Licences and versions read from Hugging Face and GitHub on November 18, 2026:

ModelSizeLicenceLanguagesVoice cloningStreamingBest for
Kokoro-82M82MApache-2.0Multiple (English plus several others)No (built-in voices)Fast enough for real timeNarration, CPU or edge
Chatterbox Multilingual V3 (Resemble AI)500MMIT23+✅ zero-shot—Cloning across languages
Qwen3-TTS (Alibaba)0.6B / 1.7BApache-2.010✅ 3-second clone (Base)✅ dual-trackMultilingual agents, voice design
VibeVoice-Realtime-0.5B (Microsoft)0.5BMITEnglish (9 experimental)No (single speaker)✅ ~300 ms first audioEnglish voice agents
Kyutai TTS 1.6B~1.8BCC-BY-4.0English, FrenchVoice conditioning✅Streaming EN/FR
Magpie TTS Multilingual (NVIDIA)364MNVIDIA Open Model License12No (5 voices)—NVIDIA stacks, commercial
Fish Audio S2 Pro4B + 400MResearch licence, non-commercial80+✅—Research and evaluation only

For hosted APIs (ElevenLabs, Cartesia, OpenAI and others) see best text-to-speech APIs; for the speech-to-text side, see best local speech-to-text models.

1. Kokoro-82M — best lightweight model

Kokoro’s 82 million parameters make it the cheapest model here to run: it works on CPU and small GPUs, and Hugging Face shows over 13 million downloads. It ships fixed voices rather than cloning, and its weights are Apache-2.0, so it is used in many commercial APIs. Install with pip install kokoro (0.9.4). The weights have not changed since April 2025, but it remains the default for audiobooks, notifications and anything needing high volume at low cost.

2. Chatterbox Multilingual V3 — best for voice cloning

Resemble AI’s Chatterbox family is MIT-licensed. Multilingual V3 (500M) improves speaker similarity, reduces hallucinated words and covers 23+ languages; a Single Language Pack adds dedicated fine-tunes for six priority languages. The English Chatterbox model adds CFG and “exaggeration” controls for expressive delivery. Install with pip install chatterbox-tts (0.1.7).

3. Qwen3-TTS — best multilingual streaming

Qwen3-TTS comes in 0.6B and 1.7B sizes under Apache-2.0, covering Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian. Base models clone a voice from a 3-second sample; CustomVoice models offer nine premium voices with natural-language style instructions (emotion, prosody). Its dual-track architecture streams audio after the first characters arrive, which suits LLM-driven agents.

4. VibeVoice-Realtime-0.5B — best for English voice agents

Microsoft’s realtime variant accepts streaming text input and produces first audible speech in about 300 ms, so an LLM can start speaking before it finishes writing. It is single-speaker and intended for English (nine other languages are experimental). The larger VibeVoice models handle multi-speaker, long-form audio such as podcasts. MIT licence.

5. Kyutai TTS, Magpie and Fish S2 Pro

Kyutai TTS (CC-BY-4.0) is a streaming English/French model built on the Moshi architecture; attribution is required. NVIDIA Magpie TTS Multilingual (364M) covers 12 languages with five voices under the NVIDIA Open Model License, which permits commercial use. Fish Audio S2 Pro has the widest language coverage (80+) and inline emotion tags, but its licence restricts it to non-commercial use.

How to choose

  1. Licence first: if the product is commercial, drop Fish S2 Pro and read Kyutai’s attribution terms.
  2. Latency: for voice agents, test streaming models inside your pipeline; see best open-source voice agent frameworks.
  3. Languages: Chatterbox (23+) and Magpie (12) beat Qwen3-TTS (10) on breadth.
  4. Hardware: Kokoro runs on CPU; models of 0.5B and larger want a GPU for real-time output.

Last verified: November 18, 2026. Licences from each model’s Hugging Face card; package versions from PyPI.

Sources