Local dictation, end to end.
Whisperstream is a dictation toolkit for Windows. Core transcription runs on your PC, and downloadable local cleanup is available on compatible hardware. Licensing, updates, downloads, and optional cloud, Agent Speak, review, feedback, or external-destination features use the internet as described in the relevant feature and privacy documentation.
The everyday pipeline: capture your voice, transcribe it on your CPU, and deliver it to many compatible focused fields.
Whisperstream transcribes your speech entirely on your computer using NVIDIA's open-weight Parakeet model running on your CPU. Your audio is never uploaded, streamed, or stored on a server, and you do not need an account. After the one-time model download, transcription works with no internet connection at all.
Dictation is push-to-talk by default: hold your hotkey, speak, and release, so nothing is ever listening in the background. The default is Right Shift, and you can remap it to any key from the Controls tab. Prefer hands-free? Switch to toggle mode, where one press starts recording and another stops it. No admin rights are needed to register the shortcut.
Whisperstream delivers transcribed text to the focused field by clipboard paste or optional simulated keystrokes. It works in many applications, including Outlook, Word, Slack, VS Code, browsers, and terminals, without an add-in. Application, security, privilege, administrator, and remote-session policies can still block or limit either method.
Whisperstream can keep a private history of completed dictations, so you can find a past transcript, replay saved audio, and copy the result again later. History is enabled by default, can be disabled, and is encrypted at rest on your device rather than synced by Whisperstream. Search by keyword, play saved audio beside its text, and delete records you no longer want.
Beyond live dictation, Whisperstream transcribes audio files you select on your device without uploading or modifying the source audio. If cloud enhancement is enabled, the resulting transcript can be sent to the provider you configured. When transcript saving is on, the finished result lands in history, ready to search, replay when audio was saved, and copy.
Whisperstream transcribes 39 languages, including English, Spanish, French, German, Chinese, Japanese, Korean, Arabic, and Hindi. NVIDIA Parakeet covers 25 languages and Qwen3 ASR adds 14 more, including Cantonese, Indonesian, Malay, Persian, Thai, Turkish, and Vietnamese. Pick your language and Whisperstream selects the on-device model, downloading a matching model when needed.
After its required speech model is downloaded, Whisperstream can record and transcribe without a connection. Updates, model downloads, license activation and periodic validation, user-configured cloud enhancement, and the separate Claude Code workflow used by Agent Speak can still use the network. Core dictation remains available when your network is down, subject to license state.
Whisperstream can lower or mute your system volume while you dictate, so music or a video does not bleed into your recording or distract you. Choose off, reduce to 20 percent, or full mute in the Audio settings. Volume returns to its previous level the moment you release the hotkey.
Optional tools that clean, shape, and tune your output. All off by default, all yours to switch on.
Whisperstream can optionally clean up your dictation: it removes filler words, fixes grammar and punctuation, and applies your spoken self-corrections (say "Friday, sorry Saturday" and it keeps just "Saturday"). Use the downloadable Local model on supported hardware, Ollama, or a cloud or custom provider you configure. It is off by default, so you choose if and where to turn it on.
Each profile is its own AI prompt, and Whisperstream switches between them by detecting the focused application locally and reducing supported browser address data to a hostname. Start from the built-in modes or write your own profile with any prompt you want: terse for your terminal, polished for email, casual for chat. Route by app or website, with a default everywhere else.
The dictionary lets you add word-for-word overrides so Whisperstream spells names, technical terms, and acronyms the way you want them. Add an entry once in the Dictionary tab and it applies to every transcription after. Overrides are case-insensitive and matched as you dictate, so you stop fixing the same word by hand.
Beyond the built-in modes, you can give the AI cleanup your own prompt with custom formatting or style instructions, and every enhanced transcription comes back shaped to match. Keep a formal tone, stay terse, use British spelling, format as bullet points, or follow your own house style. Your prompt runs on a local model or any OpenAI-compatible provider, automatically, so dictated text lands already styled the way you want.
What Whisperstream runs on, and why the install is one you can trust.
The Whisperstream installer is code-signed with a Microsoft-trusted Authenticode certificate under the publisher name Lanreal Technologies Inc. The signature confirms the download came from us and was not tampered with, and gives IT teams the exact publisher name they need to allowlist Whisperstream on managed machines. SmartScreen may still warn while a new binary or publisher builds reputation.
Lockdown Mode hardens Whisperstream for high-sensitivity work. Transcription stays on your PC, AI cleanup is limited to an on-device model, and Whisperstream cloud enhancement and diagnostics are turned off. Text still goes to the destination application you choose, and Agent Speak remains a separate Claude Code path. History can be encrypted behind a password and a content-free local access log records access events.
Core transcription runs entirely on your CPU, so you do not need a dedicated graphics card. A modern processor delivers near-instant results; a GPU is only used for the optional AI cleanup feature. The app works on 8 GB of RAM, with 16 GB recommended.
Free to try. No account.