Backend compatibility

Choose a backend profile for the API your server exposes. Choose a model profile separately for the languages and controls of its loaded model. The server can run on your computer, a LAN server, a VPS, or a hosted service.

Compare backendsBrowse the visual directory by transcription, cleanup, and speech capabilities.

Choose model behaviorFind supported languages, recognition hints, cleanup, and voice controls.

Version compatibility

Your installed release may have fewer options than these guides. Check each guide’s server version and model requirements before configuring a connection.

Generic OpenAI-compatibleConnect OpenAI-compatible services for transcription, cleanup, or text to speech.SpeachesUse one speech server for dictation, audio files, and text to speech.NeMo-Speech.cppTranscribe with Parakeet or Nemotron and generate speech with MagpieTTS.whisper.cppTranscribe microphone recordings and audio files with whisper.cpp.vLLMConnect speech recognition and text cleanup models through vLLM.llama.cppPolish completed transcripts with a text model running in llama.cpp.vLLM-OmniGenerate speech with Qwen3-TTS preset voices, languages, and style instructions.Kokoro-FastAPIGenerate spoken audio with Kokoro-FastAPI.Planned profilesUpcoming integrations and alternatives.

Run a backend

Let Freehand manage a local installation

Start with Local runtimes for supported NeMo, whisper.cpp, and llama.cpp models and platforms, including optional MagpieTTS speech alongside NeMo transcription. Downloads are explicit.

Use the recipes below when you manage the service yourself, need other models or runtime settings, or run the server on another machine.

Server	Launch instructions	Freehand operation
whisper.cpp	Native Windows or Docker	Transcription with a server-loaded model
vLLM	Windows + Docker, NVIDIA GPU	Qwen3-ASR transcription or S1-mini cleanup
Speaches	Docker GPU or CPU	Transcription; optional speech playback
Kokoro-FastAPI	Docker CPU or GPU	Speech playback with voice discovery
llama.cpp	Native Windows server	S1-mini or compatible custom cleanup
NeMo-Speech.cpp	Runtime installation	Completed/live transcription and MagpieTTS speech
vLLM-Omni	Qwen3-TTS serving guide	Text to speech with preset voices
Generic OpenAI-compatible	Connect an existing endpoint	The operations implemented by your server

Before running a local recipe

• Shell and containers: examples use PowerShell and loopback ports. Docker recipes require Linux containers; NVIDIA examples need working WSL2 GPU access. Follow Docker’s GPU setup if docker run --gpus ... cannot access the card.
• Downloads and storage: first use downloads the image and selected model. Stopping a container keeps weights in its recipe’s folder or named volume.
• Resources and shutdown: leave memory for every concurrent model or stop unused servers. Recipes include stop/restart commands or link upstream setup.

For a remote or hosted deployment, use the same profile with the administrator’s HTTPS base URL, model ID, and authentication. Loopback recipes serve a client on the same computer. See connection topologies for separate transcription, cleanup, and speech endpoints.

Check your connection

Action	What it checks
Check connection	Health or model metadata
Refresh voices	Voice metadata on supported speech servers
Short recording or voice preview	Runs only the model you explicitly selected so you can review its output

Health and discovery never generate audio or run transcription.

Choose profiles per operation

Where	Configure
Connections on the activity rail	Endpoint, authentication, backend profile, and allowed uses
Workflow Settings cog	Active connection, model, model profile, voices, and model options
Local runtimes	Installation and loaded models for managed connections

Voice transcription, Audio-file transcription, Cleanup, and Text to speech select their own connection. Tasks using one managed runtime share its selected model and get the matching profile automatically. Manual model options are remembered per connection and model.

Beyond the current contract

Transcription context and temperature depend on the selected backend and model profile. Vocabulary supplies recognition terms through the supported request fields. See the language guide for language selection.

Voice-style instructions are available with Qwen3-TTS on vLLM-Omni. Translation, timestamps, diarization, reference-audio inputs, and progressive speech playback are not available. A server may offer these features without Freehand exposing them.

Use the protocol reference for exact request fields and safety limits.