Open-source TTS has quietly closed the gap with commercial voice APIs. In 2026 you can run a text-to-speech model on a laptop CPU that sounds close to what ElevenLabs charged premium rates for two years ago — with no per-character billing, no rate limits, and no audio leaving your infrastructure. The hard part is choosing the right model, because the leading options make very different trade-offs.
This guide compares the four open-source TTS models developers actually deploy — Kokoro-82M, XTTS v2, Piper, and Chatterbox — across voice quality, speed, licensing, hardware requirements, and voice cloning, so you can pick the right one for your project.
Why Open-Source TTS Beats Paid APIs for Many Projects
Commercial voice APIs are excellent, but they bill per character, add network latency to every request, and require sending your text (and sometimes your users’ data) to a third party. Self-hosted open-source TTS flips all three: generation is effectively free after setup, latency is bounded by your own hardware, and everything stays on-premise — a hard requirement in healthcare, finance, and many enterprise deployments.
The pattern mirrors what happened with speech recognition, where Whisper made local transcription the default choice (see our Speech-to-Text API 2026 comparison). TTS is now at the same inflection point, and if you are building end-to-end voice pipelines, our guide to AI voice agent frameworks covers the orchestration side.
Kokoro-82M: The Best Default Open-Source TTS
Kokoro-82M is the model most teams should try first. At just 82 million parameters — small enough to run on a phone — it produces remarkably natural narration from the StyleTTS 2 lineage, and it is released under Apache 2.0, so commercial use is unrestricted.
- Voices and languages: 54 fixed voices across 8 languages
- Hardware: runs faster than real time on CPU; ~2–3 GB VRAM if you use a GPU
- Cost: hosted providers serve it for under $1 per million characters
- Limitation: no voice cloning — you pick from the built-in voices
Kokoro is built for narration: articles, product videos, accessibility features, and voice agent responses. If you do not need cloning or heavy emotional range, it is the best quality-per-watt in open-source TTS today.
XTTS v2: Best Voice Cloning From a Six-Second Sample
XTTS v2 remains the reference for zero-shot voice cloning: give it roughly six seconds of reference audio and it reproduces the speaker’s voice across 17 languages. Nothing else open-source clones as convincingly from so little data.
There are two important caveats. First, latency: time-to-first-audio sits around 600 milliseconds, which stacks on top of your speech recognition and LLM latency and usually pushes a live conversation past the comfortable threshold. XTTS is best for asynchronous work — audiobooks, dubbing, personalized video clips. Second, licensing: XTTS v2 ships under the Coqui Public Model License, which prohibits commercial use. For commercial cloning you will need Chatterbox or a paid API.
Piper: Lightweight TTS for Raspberry Pi and Edge Devices
Piper is the minimalist’s choice. Per-voice model files are tiny, it needs no GPU at all, and it synthesizes in real time on hardware as modest as a Raspberry Pi 4 — which is why it powers so many Home Assistant and embedded projects.
The trade-off is quality: Piper’s output is clear and intelligible but audibly synthetic next to Kokoro. For status announcements, smart-home responses, kiosks, and offline accessibility tools, that trade is usually worth it. For customer-facing audio, it is not.
Chatterbox: The Wildcard With Emotion Control
Resemble AI’s Chatterbox is the most interesting recent entrant: a 0.5B-parameter, MIT-licensed model that combines voice cloning with a unique emotion-exaggeration control. In Resemble’s blind listening study, a majority of listeners preferred Chatterbox output over ElevenLabs — a first for an open-weight model.
- License: MIT — commercial use, self-hosting, and modification all permitted
- Cloning: yes, with adjustable emotion intensity
- Extras: built-in PerTh watermarking for responsible deployment
- Hardware: larger than Kokoro — plan for a GPU in production
Open-Source TTS Comparison at a Glance
| Model | Params | License | Cloning | Best For |
|---|---|---|---|---|
| Kokoro-82M | 82M | Apache 2.0 | No | Narration, agents, commercial apps |
| XTTS v2 | ~460M | CPML (non-commercial) | Yes | Audiobooks, dubbing, research |
| Piper | Tiny per-voice | MIT | No | Raspberry Pi, edge, smart home |
| Chatterbox | 0.5B | MIT | Yes + emotion | Commercial cloning, expressive audio |
Which Open-Source TTS Model Should You Choose?
- Default choice / commercial narration: Kokoro-82M — best quality per compute, permissive license
- Voice cloning for commercial products: Chatterbox — MIT license plus emotion control
- Cloning for personal or research projects: XTTS v2 — best zero-shot clone quality
- Edge devices and offline setups: Piper — real time on a Raspberry Pi
Frequently Asked Questions
Is open-source TTS good enough to replace ElevenLabs?
For narration and agent responses, yes — Kokoro and Chatterbox are within striking distance of commercial quality, and Chatterbox has even won blind listening tests. Paid APIs still lead on very expressive multi-speaker content and turnkey convenience.
Can I use XTTS v2 in a commercial product?
No. XTTS v2 is released under the Coqui Public Model License, which restricts it to non-commercial use. For commercial voice cloning, use MIT-licensed Chatterbox; for commercial narration without cloning, use Apache-licensed Kokoro.
What hardware do I need to run Kokoro locally?
A modern CPU is enough — Kokoro generates faster than real time without a GPU. With a GPU, 2–3 GB of VRAM covers it, which means almost any card from the last several years works.
Which open-source TTS is best for real-time voice agents?
Kokoro, because its time-to-first-audio is low enough to keep total pipeline latency conversational. XTTS v2’s ~600 ms head latency is usually too slow once you add speech recognition and LLM inference on top.
Conclusion: Start With Kokoro, Reach for Chatterbox
The open-source TTS landscape in 2026 has a clear shape: Kokoro-82M for high-quality commercial narration on minimal hardware, Chatterbox when you need cloning and emotion under a permissive license, XTTS v2 for non-commercial cloning experiments, and Piper for the edge. Spin up Kokoro first — it takes minutes — and upgrade only when your use case demands cloning or more expressive range.
Building a full voice pipeline? Pair your TTS pick with the right transcription engine from our speech-to-text comparison and subscribe to NewsifyAll for weekly hands-on AI tooling guides.
