Whisp

A lightweight desktop speech-to-text tool that supports both OpenAI's transcription API and local Whisper models. Whisp provides a simple interface for converting speech to text with minimal resource overhead.

Overview

Whisp offers an unobtrusive and customizable way to transcribe your voice into text. It runs in your system tray and is globally available. Press the hotkey to start recording, speak as long as you want, then press the hotkey again to transcribe and automatically paste into any focused input field.

Design principles:

• Reliable: Built to be stable and handle errors gracefully. Resilient in the face of errors. Retry and recovery.

• Lightweight: Resource-efficient, minimal system impact, simple.

Installation

Standard Installation (OpenAI API)

cargo install whisp
whisp
# if cargo bin is not in your path
~/.cargo/bin/whisp

With Local Whisper Support

To use local Whisper models (no internet or API key required), build with the local-whisper feature:

cargo install whisp --features local-whisper

Requirements for local whisper:

• CMake must be installed (for building whisper.cpp)
• Linux: ALSA, GTK3, and X11 libraries are required. A shell.nix is provided for Nix users that includes all dependencies.

The local backend uses Metal GPU acceleration on Mac for fast transcription, with automatic CoreML support for ~3x faster encoding via the Apple Neural Engine.

Configuration

Configuration is managed through a whisp.toml file located in your system's configuration directory. The whisp tray menu has an option to copy the configuration file path to the clipboard.

Example: OpenAI Backend

backend = "openai"
hotkey = "shift+super+Semicolon"
openai_key = "your-api-key"
language = "en"
model = "gpt-4o-mini-transcribe"
restore_clipboard = true
auto_paste = true

Example: Local Whisper Backend

backend = "local"
local_model = "large-v3-turbo-q8_0"
hotkey = "shift+super+Semicolon"
language = "en"

On first launch with local backend, the model will be automatically downloaded (~834 MiB for large-v3-turbo-q8_0). The large-v3-turbo model offers excellent transcription quality and runs fast on modern hardware, especially MacBooks with Metal/CoreML acceleration.

Quick start: If you just want to test whisp, use a smaller model like tiny.en-q8_0 (~42 MiB) or base.en-q8_0 (~78 MiB) for faster initial setup.

Configuration Options

Available Local Models

Models are downloaded from ggerganov/whisper.cpp on HuggingFace. Model names must match exactly as shown below.

Usage

1. Run whisp - it will appear in your system tray
2. Press the hotkey to start recording
3. Speak as long as you want
4. Press the hotkey again to transcribe and paste

Common Use Cases

• AI Coding Agents: Voice dictate prompts to tools like Claude Code. Much faster than typing.

• Messaging: Quickly respond to messages in chat applications like Discord or Slack.

• Document Writing: Speak freely to draft large amounts of text quickly.

CoreML Acceleration (macOS)

On macOS, the local backend automatically uses CoreML to run the encoder on Apple's Neural Engine, providing ~3x faster encoding performance. This is enabled by default.

On first use, whisp downloads a pre-built CoreML encoder model from HuggingFace (15-1200 MiB depending on model size). The first transcription after launch may be slow as macOS compiles the model for the Neural Engine, but subsequent transcriptions are fast. Since whisp is designed to run in the background indefinitely, this one-time warmup cost is negligible in practice.

CoreML encoders work with all GGML models, including quantized variants. The encoder runs on the Neural Engine while the decoder uses the GGML model on CPU/Metal.

To disable CoreML and use Metal-only acceleration:

coreml = false

License

Whisp is licensed under the MIT license.