Local demo account — stored only in this browser. No server, no password leaves your device. Use any email to try it.
Hosted version uses Supabase Auth. This local build uses localStorage so you can test offline.
On phone, Flow uses tap-to-toggle Record Mode (tap once to start, tap again to stop) — hold-to-talk is unreliable on mobile. Browser speech is NOT relied on for iPhone at all.
Honest limits: System audio on the web needs you to share a screen/tab and tick “Share audio” — it works on desktop Chrome/Edge, it does NOT work on iPhone Safari, and on Android it is limited / often mic-only. A web page cannot silently capture other apps’ audio on a phone — that needs a native app. If you pick System on an unsupported browser, Flow will explain and offer Mic + Upload instead.
Record with your phone’s own Voice Recorder / Voice Memos app, then upload the MP3 / M4A / WAV here. It transcribes with your selected provider (OpenAI / Hugging Face) below — even if live mic capture fails completely.
Web apps can’t type directly into other apps — Copy & paste outside is the web way. System-wide typing needs the native Flow keyboard app (future).
Every tap, mic request, recording and provider reply is logged here in plain words. If a button seems to do nothing, read the last line here.
The Wispr Flow ideology: setup once, then one button forever. Your cursor is the destination.
Pick Email, Slack, Doc or Code below.
Or tap for hands-free, or hold Space.
Auto-Edit removes fillers and pastes at your cursor.
Target: Email
Target: Slack
Target: Google Doc
Target: Code
Auto-Edit removes “um / uh / like / you know”, false starts and repetitions, adds punctuation & paragraph breaks, expands snippets, and applies your Dictionary spelling. Try adding “Zynthwave” in Dictionary, then dictate it.
openai/whisper-large-v3 via Inference Providers. HF serverless ASR does not reliably expose language / task controls, so language is auto-detected by the model — the Language picker above applies to Browser speech and OpenAI Whisper, not reliably to HF. HF cleanup / summaries use an open chat model (selectable below).Chrome's built-in SpeechRecognition can report “not-allowed” even when your mic permission is on — especially inside an app view, an iframe, or from file://. FlowScribe now checks the mic on its own first. If the mic works but the speech service is blocked, Flow switches automatically to Record Mode: it records you with MediaRecorder and transcribes with your chosen cloud provider (OpenAI Whisper or Hugging Face whisper-large-v3) if its key is saved.
All keys on this page are stored only in your browser (localStorage), never sent anywhere except api.openai.com / router.huggingface.co — each key is only sent to its own provider.
Create your OpenAI API key →
Create a key, copy it, paste it here, then Save & Test.
Create your Hugging Face token →
Create a fine-grained token with Inference Providers permission, copy it, paste it here, then Save & Test.
Type & clean instead always works without any mic: type or paste into Raw transcript above, then press Clean & Paste to Target. No mic, no key, no speech service needed.
Every dictation, recording, upload and pop-up session autosaves the moment you stop — raw + cleaned text, destination, provider, duration and word count — in this browser (localStorage), so nothing is lost on reload or when you close the pop-up. Tap a saved item to reopen it in the editor, copy it, export it, delete it, or continue dictating into it.
Upload audio, edit, summarise and export — all locally.
“Open in editor” focuses the large text area above for editing. ChatGPT opens in a new tab at chatgpt.com — the transcript (plus the instruction you pick) is copied first, so you paste it there with Ctrl+V. FlowScribe cannot type directly inside ChatGPT.
Command Mode is Pro in the hosted plan — in this local build, toggle Pro free in Settings → Billing to test it. History:
With OpenAI or Hugging Face selected (and its key saved), Summarise / Command Mode below call that provider’s chat model. With Browser selected, or if the cloud call fails, the local rules-based summary is used and the failure is shown — never hidden.
Record, label speakers, get a summary, and ask questions about the meeting.
Same source as Dictate. System = share the meeting tab (desktop Chrome/Edge). It does NOT work on iPhone Safari and is limited on Android — use Mic + Upload there.
Unlimited speakers: use “+ Add speaker” for Speaker 3, Speaker 4 and beyond. Click a speaker name to rename it, pick its colour, or delete it. Every transcript segment below has a speaker dropdown with all speakers plus “Add new speaker…”. Meetings autosave as you speak; “Save meeting” stores title, date, speakers, timed segments, summary and action items in this browser so you can reopen them later.
Reopen, continue, edit, export or delete a saved meeting.
Words FlowScribe should always spell your way. Auto-Edit uses these live.
Say the trigger while dictating and it expands to the full text.
Words today
Words this week
Average speed • 0 min saved vs typing
0 / 30 minutes
0 / 2,000 words
Off by default. Voice wake listens only while Hands-free mode is on (the Hands-free checkbox in Dictate) — it never listens at other times. When it hears a wake phrase it opens the Flow Pop-up and starts listening there. If voice wake is off, the floating Flow bubble, the Flow button, holding Space, or your hotkey above is the wake.
Free: 30 min / 2,000 words per week.
Pro: $12/mo billed annually • $15/mo monthly — unlimited dictation, Command Mode, Notetaker AI.
Hosted version uses Stripe Checkout + Portal. Locally this is just a toggle so you can test Pro features.
Same choice as in Dictate — changing it in either place updates both. Browser = Chrome speech, free, no key. OpenAI / Hugging Face = Record Mode recordings + chat model for cleanup, summaries & Command Mode.
Create your OpenAI API key →
Create a key, copy it, paste it here, then Save & Test.
POST https://api.openai.com/v1/audio/transcriptions (model whisper-1) when you record in Record Mode.Create your Hugging Face token →
Create a fine-grained token with Inference Providers permission, copy it, paste it here, then Save & Test.
POST https://router.huggingface.co/hf-inference/models/openai/whisper-large-v3 (model openai/whisper-large-v3, audio sent as base64, return_timestamps: true where supported; language auto-detected by the model). Chat: OpenAI-compatible POST https://router.huggingface.co/v1/chat/completions.Hugging Face is not free for this any more. Since 7 Oct 2026, free HF accounts get no monthly Inference Providers credit — free users must buy credits, or go PRO at $9/mo (which includes $2/mo of credit). Usage beyond included credit is billed per provider / model.
OpenAI Whisper: $0.006/min (about $0.36/hour) for whisper-1, plus chat-model charges for cleanup / summaries if you use OpenAI for those. Browser speech: free, no key, but it is the flaky one in app views / file://. Check current prices on each provider’s site before heavy use.
Hosted integrations activate with your own keys in the hosted version: OpenAI Whisper (POST /v1/audio/transcriptions), Supabase Auth + Postgres (users, transcriptions tables) + Storage bucket audio_files, Stripe webhooks (checkout.session.completed, customer.subscription.updated/deleted). Locally: browser SpeechRecognition is tried first when Browser is selected; when Chrome blocks it, or when OpenAI / Hugging Face is selected, MediaRecorder + that provider’s key above is used, honestly labelled on every transcript.
Pick a destination — everything you dictate is imprinted there live, word by word, cleaned.
“Open in main app” jumps straight to this transcript in Saved transcripts / the editor, highlighted. Everything dictated here is already saved to the main app.
ChatGPT opens in a new tab at chatgpt.com — your transcript is copied first, so you paste it there (Ctrl+V / long-press → Paste). FlowScribe cannot type directly inside ChatGPT.
Standalone bubble app. Everything you dictate here is saved instantly to the main FlowScribe app — same browser store, live-synced if the main app is open in another tab or window.
Hold to talk, or tap to start / tap again to stop. Release saves automatically.
ChatGPT opens in a new tab — your transcript is copied first, so you paste it there (Ctrl+V). FlowScribe cannot type directly inside ChatGPT.
Tip: keep this small window beside your work. Use “Open in main app →” to jump straight to what you just dictated, highlighted in the editor.
Pick what ChatGPT should do. The chosen instruction is copied together with your transcript as one paste. ChatGPT then opens in a new tab — paste there with Ctrl+V (or long-press → Paste). FlowScribe cannot inject text directly into ChatGPT.
1. Switch to your other app (WhatsApp, Notes, Gmail…)
2. Tap where you want the text
3. Paste (long-press → Paste on phone, Ctrl+V on desktop)
Web apps can’t type directly into other apps — Copy & paste outside is the web way. System-wide typing needs the native Flow keyboard app (future).