Self-Hosting voice services (TTS, ASR, Wake Word)
TL;DR https://codeberg.org/cryptomilk/crane-wyoming
Where I started
I use Home Assistant, and for text-to-speech (TTS) I’ve been running Piper through Wyoming Piper.
Piper is a fast, local neural TTS engine originally built for the Rhasspy project and now maintained by the Open Home Foundation. It’s designed to run entirely offline, even on modest hardware like a Raspberry Pi.
Wyoming is the open protocol Home Assistant uses to talk to voice components like TTS, speech-to-text, wake word, voice activity detection (VAD, which decides when someone has started or stopped speaking) over the network, so any service that speaks Wyoming can be plugged in as a satellite. Wyoming Piper just wraps Piper so it can be served this way.
Both work well and I have no complaints about reliability. My issue is quality: the German voices aren’t great. Piper depends on open datasets for training, and good open German speech data is scarce, so the German models lag behind the English ones. I also run TTS locally on my desktop for event reminders, so voice quality matters to me beyond just Home Assistant.
Looking for something better
I wanted better output quality, so I started looking at alternatives and found Crane, a Rust inference
framework built on Candle.
An “inference model” is a trained neural network used to actually produce output like text, speech, an image, rather than to learn from data (that’s “training”). An “inference framework” is the software that loads such a model and runs it efficiently: managing GPU/CPU memory, batching requests,
and exposing an API around it. Piper and Crane are both inference frameworks.
Crane already had Qwen3-TTS support, and its Serena voice’s German output sounded noticeably better. I also wanted to try Voxtral-4B-TTS-2603, Mistral’s open-weight TTS model, so I added support for it. Voxtral TTS produces expressive, natural-sounding speech across 9 languages including German, with low time-to-first-audio and streaming support. It is a good fit for a voice assistant that needs to start speaking quickly.
Adding Wyoming support
Once Voxtral was working in Crane, I built crane-wyoming, a standalone Wyoming protocol server, so Home Assistant could use these models as its TTS service. To make that possible, I added the Tts trait and the surrounding TTS abstractions to Crane, since there was no stable interface for driving a TTS model on its own, separate from Crane’s full inference engine (tokenizer/LLM/VLM machinery). Those abstractions have since been merged upstream: lucasjinreal/Crane#44.
crane-wyoming depends on Crane only for that Tts trait and the concrete model types it needs to construct, not Crane’s engine crate. So it carries its own small TTS-only model runtime (one dedicated worker thread per loaded model) and its own on-disk response cache.
The project grew into a small Cargo workspace. Besides the Wyoming server
itself, it now has cw-say, a standalone CLI client for scripting.
It also has sd_crane_wyoming, an output module for speech-dispatcher. speech-dispatcher is the common Linux TTS abstraction layer that screen readers like Orca, and other accessibility tooling, talk to. It launches output modules as subprocesses and speaks to them over stdin/stdout, using its own line-oriented, SMTP-style protocol. sd_crane_wyoming translates that into Wyoming requests against a running crane-wyoming server. That way, the same server process and cache serving Home Assistant can also serve the desktop. After registering it in speechd.conf, spd-say -o crane "..." works. So does anything else built on speech-dispatcher, like Firefox’s “Read Aloud” or Orca itself. All of it gets the same voice quality as Home Assistant, without running a second TTS backend.
What’s next
- Speech-to-text. I’ve added Qwen3-ASR support for utomatic speech recognition (ASR) into Crane. This needs to be wired in crane-wyoming next. Also Voxtral-Mini-4B-Realtime-2602 is interesting.
- VAD. Crane already has a Silero VAD implementation. Adding it for STT is straight forward.
- Wake word. Once VAD is in place, add Open Wake Word support or similar.
With all of that implemented, you’d have a complete self-hosted Wyoming voice stack with no cloud dependency.
Current limitations
The catch is that you need a GPU to run it well.
If you only need TTS for occasional things like reminders, short announcements, running it on CPU with caching is enough, since repeated phrases just get served from cache instead of resynthesized.
All of this is for advanced users and hackers right now. There’s no polished packaging yet. Systemd units exist for both system and user services, including socket activation, but you still have to build from source.
However testing and feedback are welcome.
