Local Voice-AI Model Database: Self-Hosted STT & TTS, License-Graded (2026)

Local large language models get all the attention, but a genuinely private AI stack also needs to hear and speak — and the speech models follow different rules. The one that decides what you can actually ship isn’t quality, it’s the license, and text-to-speech in particular is a licensing minefield where the best-sounding models are the ones you legally can’t use in a product. This database scores 21 self-hostable speech-to-text and text-to-speech models on license commercial-safety first, then size, languages, speed by hardware, streaming, and whether they run on a Raspberry Pi — so you can find the model you can both afford to run and legally deploy.

Speech-to-text is easy; text-to-speech is a minefield

On the transcription side, you’re spoiled for commercial-safe choice. The entire Whisper family is MIT-licensed, and almost nobody runs OpenAI’s original Python — the real workhorses are whisper.cpp (a single C/C++ binary that runs down to a Raspberry Pi) and faster-whisper (four times faster on a GPU), both MIT. NVIDIA’s Parakeet edges out Whisper on English accuracy-per-watt under a permissive CC-BY licence, and Kyutai’s streaming model happens to be natively English-and-French — the exact bilingual pair a Canadian tool needs. Text-to-speech is where it gets treacherous: the two best open voice cloners, XTTS-v2 and Fish Speech, carry non-commercial licences, and in XTTS-v2’s case the company that could have sold you a commercial licence no longer exists. If you need commercial-safe synthesis, the answers are Kokoro-82M and Chatterbox (permissive, and both punch above their weight), with Piper for anything that has to run on tiny hardware.

What powers DCENT_Voice

A local, private voice-to-text tool like our own DCENT_Voice realistically ships the whisper.cpp / faster-whisper family: permissively licensed end to end (the Whisper weights are MIT too), no Python runtime to bundle, quantized GGML models that fit on modest hardware, and a clean CPU-only path all the way down to Pi-class devices, with a GPU fast-path for anyone who has VRAM. That’s not a marketing choice — it’s what the licence and hardware columns in the table above force if you want something you can actually distribute. For the models that generate text rather than transcribe it, see the companion local LLM model database; to size the hardware, the VRAM calculator; and to check whether any of it phones home, the telemetry & air-gap audit. Everything ties back to the Local LLM Canada hub.