talx

context

I was using Wispr Flow daily on the free Pro trial. It genuinely changed how I wrote. Once the trial ended I was down to 2,000 words a week, which is nothing if you actually dictate every day. The paid tier is $12–15 a month for the same idea I could rebuild on a free API.

Talx is that rebuild. Hold a hotkey, speak, and polished text gets pasted wherever the cursor is. Open source, MIT.

what i built

A floating macOS widget that lives system-wide. Hold the hotkey (default ⌥ Space), speak; release, and the polished text gets pasted into whatever app is in front. Works in any text field, any app.

Two ASR providers: Groq Whisper for free batch transcription, Soniox for real-time streaming. Two LLM providers for polishing: Groq (free) or OpenAI. Model selectable per provider. Session history and a small usage dashboard, all in local SQLite.

engineering choices

Tauri v2 over Electron. Smaller binary, real native shell, Rust backend for the parts that matter: audio capture via cpal, hotkey handling, IPC.

Bring-your-own keys, stored in local SQLite. Nothing routes through a server I run. No telemetry, no analytics. Audio goes straight from the user's machine to the provider they picked.

Honest pipeline: record → ASR → LLM polish → paste. Polish lives behind a separate provider call so the raw transcript stays inspectable and failure modes don't compound.

why it's free

Groq's free tier gives roughly 8 hours of audio transcription per day plus 1,000 LLM polishing calls. That works out to about 72,000 dictated words daily — roughly 250× Wispr Flow's free weekly word budget. The only real ceiling is a per-minute rate limit that normal dictation never hits.

outcome

Open source on GitHub. Used daily on my own machine. macOS-first today; Windows and Linux are on the roadmap.