Wispr Flow

Captures audio input from the user's microphone, processes it through speech-to-text conversion (likely using cloud-based ASR like Whisper API or similar), and injects the resulting text directly into the active application's input field via OS-level keyboard event simulation. This works across any application (browsers, IDEs, email clients, etc.) without requiring native integration, by hooking into the operating system's input pipeline rather than relying on application-specific APIs.

Processes continuous audio stream from microphone through a speech-to-text engine (architecture suggests cloud-based ASR, possibly Whisper or similar), applying automatic formatting rules to convert raw transcription into properly punctuated, capitalized prose. The system likely maintains a buffer of recent audio to handle edge cases like sentence boundaries and applies post-processing rules for common patterns (capitalization after periods, removing filler words, etc.).

Detects the currently active application window and potentially routes voice input differently based on application type (e.g., IDE vs email client vs browser). While not explicitly documented, this capability likely uses OS window focus detection and application identification to determine whether to treat input as prose, code, or structured data. The system may maintain a registry of application profiles that define how text should be formatted or injected.

Implements efficient audio capture from the system microphone with minimal buffering and streaming architecture to send audio chunks to a remote speech recognition service. The system likely uses a ring buffer or chunked streaming approach to minimize latency between speech end and text output, with potential local audio preprocessing (gain normalization, silence detection) to optimize cloud ASR performance and reduce bandwidth usage.

Provides a global hotkey (likely configurable) that activates voice dictation from anywhere on the system, independent of application focus. The system manages voice session lifecycle — detecting hotkey press, starting audio capture, detecting end of speech (via silence timeout or explicit hotkey release), and injecting text. This requires a system-level input hook that monitors keyboard events even when the application is not in focus.

Injects transcribed text into the active application using OS-appropriate input methods — simulating keyboard events on Windows/macOS, adapting to different input field types (text areas, code editors, rich text fields). The system likely detects the input field type and adjusts injection strategy accordingly (e.g., handling special characters differently in code editors vs prose editors, respecting undo/redo stacks).