Self hosted text to speech: open models and licences

Self hosted text to speech means running an open TTS model such as Piper, Kokoro, Indic Parler-TTS or Chatterbox on your own CPU or GPU, so no text or audio leaves your network. The quality is good enough for many products, but the licence matters as much as the voice: XTTS-v2 and the F5-TTS base models forbid commercial use.

These notes are from our voice AI lab, written for developers who need text to speech in a product, a phone agent or an internal tool.

Which open TTS models can you run yourself?

The models below are the ones worth testing in late 2026. The licence column is what each repository or model card says on 27 October 2026.

What does each licence let you do?

• Piper. The original rhasspy/piper repository (MIT) was archived on 6 October 2025, and development moved to OHF-Voice/piper1-gpl under GPL-3.0. If you ship Piper inside software you distribute, GPL terms apply to that distribution. The voices are separate: the project says each voice's model card holds its licence and some are restrictive, so check every voice you use.
• Kokoro, Indic Parler-TTS, Chatterbox, IndicF5. Permissive licences (Apache 2.0 or MIT) that allow commercial use. IndicF5 is gated: you agree to share your Hugging Face details before downloading.
• XTTS-v2. The Coqui Public Model License lets a company use the model only for testing, evaluation or non-commercial research. A paid product is out.
• F5-TTS. The code is MIT, but the pretrained models are CC-BY-NC because of the Emilia training data. Fine for a lab, not for a paid service unless you train your own weights on data you may use.

Which models speak Tamil, Hindi or Arabic?

For Tamil, the realistic open choices are Indic Parler-TTS and IndicF5, both from AI4Bharat. Indic Parler-TTS covers 21 languages, has 69 named speakers, and lets you describe the voice in a caption (pitch, pace, expressiveness, "very clear audio"). IndicF5 copies the voice of a reference clip.

For Hindi, you have the widest choice: Piper, Kokoro, Indic Parler-TTS, IndicF5, Chatterbox and XTTS-v2.

For Arabic, Chatterbox Multilingual (MIT) is the permissive option. Piper has an Arabic voice from Jordan, and XTTS-v2 speaks Arabic but cannot be used commercially. Test Gulf dialects with native speakers before you commit; a Modern Standard Arabic voice may sound formal on a support line.

Code switching is the hard case in India: a caller in Chennai may mix Tamil and English in one sentence. Our voice AI lab tests exactly this, because a model can read each language well and still stumble on the switch.

GPU or CPU, and is it fast enough for phone calls?

Size is a rough guide. Kokoro has 82 million parameters, Chatterbox Multilingual about 500 million, and Indic Parler-TTS about 0.9 billion. Piper runs on CPU, with GPU as an option (--cuda with onnxruntime-gpu). Resemble AI says Chatterbox Nano, an English-only 110 million parameter model, runs three times faster than real time on 8 CPU cores. We have not benchmarked these models ourselves, so treat vendor figures as a starting point.

For a live phone call, the number that matters is time to first audio: how long after the language model's first words the caller hears speech. A caller expects a reply within a second or two or starts talking over the agent, and speech-to-text and the language model take their share of that time first. So:

1. Stream text into the TTS model sentence by sentence, not the whole reply at once.
2. Measure time to first audio on your hardware, with real sentences, not total generation time.
3. Test on a real phone line, which carries lower quality audio than the model produces.
4. Keep a GPU free for TTS on a busy line, since several calls run at once.

For WhatsApp voice notes, none of this matters much: a few seconds of delay is fine.

Voice cloning needs consent

Cloning a voice from a short clip is now easy, which is why it needs rules. IndicF5's terms say to clone only voices for which you have explicit permission. Chatterbox puts an inaudible PerTh watermark in every file it generates, so its output can be identified later.

Get written consent from anyone whose voice you clone, say what it will be used for, and let them withdraw it. Tell callers they are speaking to an AI. In the United States, the FCC ruled in February 2024 that AI-generated voices, including cloned ones, count as "artificial" voices under the Telephone Consumer Protection Act, so automated calls with them need the called party's prior express consent. This is general information, not legal advice.

Where to start

For a commercial product in Indian languages, test Indic Parler-TTS first. For English or Hindi on modest hardware, try Kokoro or Piper. Use Chatterbox if you need Arabic or cloning under a permissive licence. Our voice AI agents combine speech-to-text, a language model with tools, and one of these voices, on your own server or through hosted speech APIs.

If you are choosing a voice for a product or a phone line and need it in Tamil, Hindi or Arabic, tell us the use case and we will help you test the shortlist on your own hardware. Book a consultation.