TTS Options for OpenVoiceOS in 2026
In late 2024 I wrote a guide to TTS options for OpenVoiceOS. Since then, the plugin I recommended most, Piper, was archived. Coqui had already shut down in early 2024, and several voices that sound far more natural became easy to run on a CPU.
This post covers what to pick, the full list of maintained plugins as of November 2026, and what was archived.
Quick picks #
| Your situation | Use this | Why |
|---|---|---|
| Raspberry Pi, on-device | phoonnx | The OVOS default. Runs Piper voices and several other model types through ONNX. |
| Most natural voice on a CPU | Kokoro on a server | Natural voices in 9 languages, faster than real time on a modern CPU. |
| Apple Silicon, or you want voice cloning | Pocket TTS | CPU-only, about 6x real time on Apple Silicon, clones a voice from a sample. |
| Best voice cloning, and you have a GPU | Audio8 on a server | Clones a voice from one short clip. Faster than real time on an NVIDIA GPU or Apple Silicon, but too slow on a CPU. |
| You don’t mind the cloud | Edge TTS | Microsoft’s Edge voices. Fast and good, but needs the internet and isn’t private. |
Disclosure: Oscillate Labs maintains the Kokoro, Pocket TTS, and Audio8 plugins. I use Kokoro every day.
Moving from Piper #
The Piper plugin (ovos-tts-plugin-piper) is archived. Piper support now lives in phoonnx, which also runs Matcha, GlowTTS, MMS, and other models. Install it with pip install phoonnx. You still need espeak-ng from your OS package manager, as before. Voice IDs use the format piper/<lang_code>-<name>.
Change the module and the voice ID in mycroft.conf. Before:
"tts": {
"module": "ovos-tts-plugin-piper",
"ovos-tts-plugin-piper": {
"voice": "alan-low"
}
}
After:
"tts": {
"module": "ovos-tts-plugin-phoonnx",
"ovos-tts-plugin-phoonnx": {
"voice": "piper/en_GB-alan-low"
}
}
If you leave voice unset, phoonnx picks the best voice for your language. phoonnx’s VOICES.md lists every voice ID. The Piper repo’s README covers local model files and remote URLs.
Run TTS on a server #
As with STT, you can run TTS on a bigger machine and point every assistant at it. Install ovos-tts-plugin-server on the assistant. My ovos-speech-servers repo has a tested Kokoro recipe today, with phoonnx and Pocket TTS planned. See
Self-Hosted Speech Servers for details.
OVOS also provides ovos-tts-server and Docker images. If you also run Home Assistant, you have two options. ha-ovos-tts, which I maintain, is a Home Assistant integration (installed through HACS/Marketplace) that uses your ovos-tts-server as a TTS engine. Or wyoming-ovos-tts bridges any OVOS TTS plugin to Home Assistant’s Wyoming protocol, since ovos-tts-server doesn’t speak it.
Maintained plugins (November 2026) #
Natural-sounding, offline #
| Plugin | Notes |
|---|---|
| phoonnx | The OVOS default and Piper’s successor. Multilingual phonemization plus Piper, Matcha, GlowTTS, MMS, and other ONNX models. |
| Kokoro | hexgrad’s 82M-parameter multilingual model. Needs espeak-ng. Sends timed visemes, so a Mark 1 mouth stays in sync with no G2P plugin. |
| Pocket TTS | Kyutai’s 100M-parameter, CPU-only model. Clones voices from audio samples. |
| Audio8 | Audio8 TTS Realistic 0.6B. One-shot voice cloning from a reference clip, with optional streaming. Needs a GPU: RTF 0.27 on an RTX 3090 and 0.54 on an M3 Max, but 1.49 on a CPU. Not for a Pi. Proof-of-concept status. |
| OmniVoice | k2-fsa’s zero-shot model for 600+ languages. Picks a voice automatically, designs one from a text description, or clones a reference clip. New in 2026 and still alpha. A heavy install (PyTorch plus a multi-gigabyte model), so run it on a server. I haven’t tried it. |
| Coqui | The Coqui company is gone, but this plugin is still maintained running off a fork. Best with a GPU. |
Lightweight and robotic #
| Plugin | Notes |
|---|---|
| eSpeak NG | Many languages, very light, robotic. |
| Mimic | The classic Mycroft engine. Light and robotic. |
| Pico | Very light. |
| MaryTTS | Connects to a MaryTTS server. The plugin is maintained, but MaryTTS itself is dated. |
Cloud #
| Plugin | Notes |
|---|---|
| Edge TTS | Microsoft Edge voices. Free to use, fast, and supports streaming. Not private. |
| Azure | Paid, high quality, fast, not private. |
| Polly | Amazon’s TTS. Paid, high quality, fast, not private. |
| Google Translate | Uses the undocumented Google Translate speech endpoint. The plugin’s own README says it’s a legal gray area that can break without notice. Avoid it for anything that matters. |
Language-specific #
| Plugin | Notes |
|---|---|
| Cotovia | Galician. |
| AhoTTS | Basque and Spanish. |
Novelty #
| Plugin | Notes |
|---|---|
| BeepSpeak | R2-D2-style beeps. Fun, but you’ll need subtitles. |
| SAM | Software Automatic Mouth, the retro 1980s voice. |
Post-processing #
OVOS also has TTS transformers that process audio after synthesis and before playback, so they work with any TTS plugin. They include sox effects and three super-resolution transformers. Those upscale low-sample-rate voices (such as Piper’s low voices) to 48 kHz, so speech sounds brighter and less muffled:
- FlashSR: ONNX, uses a GPU if one is available, otherwise runs on a CPU. The light option.
- AudioSR: pure ONNX, with no PyTorch at runtime. If the upscale fails, you get the original audio.
- NovaSR: PyTorch. Its own README recommends FlashSR as the lighter choice.
I haven’t tried them. If your TTS already outputs 48 kHz audio, they leave it untouched, so there’s no reason to install them.
Archived since 2024 #
These plugins are now archived. Don’t start new setups with them.
- Piper: moved into phoonnx (see above).
- Mimic 2, Mimic 3, and the Mimic 3 server: use phoonnx or Kokoro.
- Nòs (Galician) and Cotovia remote: use Cotovia.
- Matxa (Catalan multi-speaker): moved into phoonnx. The same model is
OpenVoiceOS/matxa-cat-multispeaker-vocos-2stage, and you pick a speaker withspeaker_id. The Matxa repo maps the old voices to speaker IDs. - Lux: a voice-cloning plugin built on ZipVoice. Its repo doesn’t name a replacement. For voice cloning, use Audio8 if you have a GPU, or Pocket TTS on a CPU.
What about Neon’s plugins? #
Neon’s Coqui and Coqui remote plugins are still maintained and run exclusively custom Neon-trained models. Most of Neon’s other TTS plugins haven’t changed since 2022. The OVOS plugins above all work on Neon.
Conclusion #
For a Raspberry Pi, phoonnx is the easy default. For the best voice, run Kokoro or Pocket TTS on a server and share it with every assistant in the house.
Like the STT list, this one will go stale. The ovos-speech-servers repo is tested weekly and will stay current.
Questions? Comments? Feedback? Let me know on the Conversational AI Community Forums or OVOS support chat on Matrix.