Wispr Flow is excellent, but at around $15 per month it adds up to $180 a year — and every word you dictate is sent to their cloud servers to be transcribed. If you already run a home server, you can replicate most of that functionality for free, with your audio never leaving your own hardware.
This guide explains the architecture, the tools involved, and what you need to set it up. A full step-by-step build guide is coming — this post covers the concepts, components, and how they fit together so you can decide if it is worth attempting.
Why Self-Host?
There are two strong reasons to run voice dictation on your own hardware:
- Cost — faster-whisper is open source and free. Once your server is running, transcription costs nothing per minute regardless of volume.
- Privacy — audio is processed entirely on your local network. Confidential emails, client conversations, financial discussions — none of it touches an external server. This matters if you operate under GDPR, work in a regulated industry, or simply prefer not to send sensitive business communication to a third-party cloud.
How Wispr Flow Works (and How to Replicate It)
Wispr Flow works in three steps that happen in under two seconds:
- Your hotkey press triggers audio recording on your PC or Mac
- The recorded audio is sent to a transcription server (their cloud)
- The transcript is returned and typed into your active app
Self-hosting replaces step two with your own server on your local network. The client on your Windows or Mac machine stays the same in concept — it just sends the audio to 192.168.1.x instead of Wispr’s servers.
The Stack
Server: faster-whisper
faster-whisper is an open-source reimplementation of OpenAI’s Whisper speech recognition model. It runs significantly faster than the original — up to four times quicker on CPU and even faster with a GPU — and produces the same quality output. It is what you would run on your home server to handle transcription.
Models range from tiny (fastest, slightly less accurate) to large-v3 (most accurate, requires more VRAM or RAM). For everyday dictation, the small or medium model gives a good balance of speed and accuracy on typical home server hardware.
API Layer: Speaches (formerly faster-whisper-server)
Speaches is a Docker container that wraps faster-whisper in an OpenAI-compatible API. This is important — it means any client written to talk to OpenAI’s transcription API will work against your local server with no modification beyond changing the endpoint URL.
You run Speaches on your home server and it listens on a local port (e.g., http://192.168.1.100:8000). Any device on your network can then submit audio and receive a transcript.
Client: a hotkey script on your Windows PC
The client side is a small script that:
- Starts recording your microphone when you hold a hotkey
- Stops recording when you release it
- Sends the audio file to your local Speaches endpoint
- Receives the transcript text
- Pastes it into the active window using a simulated keyboard input
This can be written in Python using standard libraries (sounddevice for recording, requests for the API call, pyperclip or pynput for pasting). The whole script is around 50–80 lines.
Hardware Requirements
The server-side transcription is the compute-heavy part. Here is what you can expect on typical home server hardware:
| Hardware | Recommended Model | Transcription Speed | Notes |
|---|---|---|---|
| Modern CPU only (e.g., Intel N100) | tiny or base | 2–4x real-time | Works, slightly slower on long dictations |
| Mid-range CPU (i5/i7) | small or medium | 4–8x real-time | Comfortable for everyday use |
| Dedicated GPU (GTX 1060+) | medium or large-v3 | 20–40x real-time | Near-instant response, Wispr-quality accuracy |
| Apple Silicon Mac mini (home server) | large-v3 | Very fast | Excellent if using Mac as home server |
For a typical spoken sentence of 10–15 words, even a CPU-only setup with the small model will return the transcript in under a second. The latency only becomes noticeable on long paragraphs.
What You Will Need
- A home server running Linux (Ubuntu, Debian, or Proxmox LXC)
- Docker installed on the server
- Python 3 on your Windows or Mac client machine
- A microphone on your client machine (built-in or USB)
- Your server and client on the same local network (or accessible via Tailscale for remote use)
Cost Comparison
| Wispr Flow Pro | Self-Hosted | |
|---|---|---|
| Monthly cost | ~$15/month | £0 (server already running) |
| Annual cost | ~$180/year | £0 |
| Audio leaves your machine | Yes (cloud) | No (local network only) |
| Works offline | No | Yes (LAN only needed) |
| Setup time | 5 minutes | 1–2 hours |
| Maintenance | None | Occasional Docker updates |
Limitations vs Wispr Flow
Self-hosting is not a perfect replacement. The honest gaps:
- No AI rewriting — Wispr Flow’s optional polish pass (rephrase, clean up) uses an LLM on top of transcription. You can add this with Ollama locally, but it adds complexity.
- Setup effort — Wispr Flow takes five minutes. The self-hosted stack takes an hour or two to configure the first time.
- No mobile support — Wispr Flow works on iPhone and Android. The self-hosted client described here is desktop-only.
- Manual updates — you manage the Docker container and model updates yourself.
Is It Worth It?
If you already run a home server, the marginal cost of adding a transcription container is essentially zero. The $180/year saving pays for itself immediately, and the privacy benefit is real for anyone handling sensitive business communication.
If you do not already have a home server, the self-hosted route only makes sense if you were planning to build one anyway — the server hardware cost would exceed several years of Wispr Flow subscriptions.
The full step-by-step build guide — covering Docker setup, Speaches configuration, and the Windows client script — is coming soon. If you want to be notified when it is published, check back or follow the site.
Related Guides
- Wispr Flow Review: AI Voice Dictation That Actually Works
- How to Build a Home Server
- How to Run Ollama on a Home Server
- Ollama Security: Safe Remote Access with Tailscale
- Plaud Note Review: AI Voice Recorder for Meetings
