Download Try Live

Local STT Setup

If you already run local LLMs with Ollama or LM Studio and a local voice with Local TTS, you can transcribe your voice locally too. The audio is processed on your machine, so nothing leaves the device and there are no API keys or per-minute costs.

Utsuwa talks to any server that exposes the OpenAI /v1/audio/transcriptions endpoint. It records from your microphone, sends the clip to that endpoint, and uses the returned text as your message.

Speaches (formerly faster-whisper-server) serves Whisper transcription through the OpenAI API and runs well on CPU.

Installation

The quickest path is Docker:

docker run -p 8000:8000 ghcr.io/speaches-ai/speaches:latest-cpu

If you have an NVIDIA GPU, use the CUDA image instead. See the project README for non-Docker installs.

This serves the API at http://localhost:8000/v1.

Connecting to Utsuwa

  1. Open Settings (gear icon)
  2. Navigate to the Character tab
  3. Open AI Services and scroll to Voice Input (STT)
  4. Under Local server, enter the base URL — leave it as http://localhost:8000/v1/ unless you changed the port
  5. Set the Model field to a model your server exposes (e.g. Systran/faster-whisper-large-v3)
  6. Click the microphone button in the chat bar and speak

A configured local server takes priority over Groq and the browser’s Web Speech API automatically — there’s no separate toggle.

Models

Speaches uses Hugging Face model IDs such as Systran/faster-whisper-large-v3 (best quality) or Systran/faster-whisper-medium (faster). The server downloads the model on first use. Check the models your server has with curl http://localhost:8000/v1/models.

faster-whisper-server and whisper.cpp

Any OpenAI-compatible transcription server works. whisper.cpp ships a server (whisper-server) that exposes the same endpoint — point Utsuwa’s Local STT base URL at it and use the model name it serves. Older faster-whisper-server installs behave like Speaches.

Custom Base URL

Running the server on a different machine or port? Enter the full URL in the Local STT base URL field. Utsuwa normalizes it to end in /v1, so http://localhost:8000, http://localhost:8000/v1, and http://localhost:8000/v1/ all work. Examples:

  • Custom port: http://localhost:9000/v1/
  • Remote machine: http://192.168.1.50:8000/v1/ (desktop app only, see below)

Desktop app vs hosted website

Local STT works best in the desktop app, where it needs no extra setup. The desktop app talks to your local server directly, with no browser origin, mixed-content, or local-network restrictions.

On the hosted website (https://www.utsuwa.ai) it can still work, but because a public HTTPS page is reaching a server on your own machine, the browser adds a few rules:

  • Same machine only. The server has to be on localhost / 127.0.0.1. A server on another machine over plain http:// is blocked by the browser as mixed content. (localhost is exempt, which is the only reason the local case works.) The remote-machine base URL above therefore works in the desktop app but not on the hosted site.
  • The server must allow the site’s origin. Your STT server needs to send CORS headers permitting https://www.utsuwa.ai. A hardened or proxied server may need the origin added explicitly.
  • Your browser may ask permission. Recent versions of Chrome treat a public site reaching localhost as a local-network request and may prompt you to allow it. Allow it if asked.

None of this applies to the desktop app. It’s the same set of rules local LLMs and Local TTS follow on the hosted site.

Troubleshooting

“Could not reach a local STT server”

The server isn’t running or isn’t reachable at the base URL. Confirm it’s up:

curl http://localhost:8000/v1/models

If that returns data but Utsuwa still can’t reach it from a browser, it’s almost certainly an origin or local-network block. On the hosted site the server has to allow the https://www.utsuwa.ai origin, and your browser may prompt to allow access to local-network devices. See Desktop app vs hosted website. None of this applies to the desktop app, which is the smoothest way to run local STT.

400 or 404 from the server

The model name isn’t valid for that server. Check the available models with curl http://localhost:8000/v1/models and set the Model field to one of them.

No transcription after speaking

Make sure your microphone is allowed for the site (browser permission prompt), and that the Model field is set. Watch the microphone button — it shows a transcribing state after you stop speaking while the server processes the clip.