Picovoice/leopard

On-device speech-to-text engine powered by deep learning

What it solves

Leopard provides a way to perform high-accuracy speech-to-text transcription directly on a device, eliminating the need to send voice data to a cloud server. This ensures user privacy and allows for offline functionality.

How it works

It is an on-device speech-to-text engine that processes audio files or microphone input locally. It uses a model-based approach where users can use a default model or a custom-trained model. The engine requires an AccessKey for authentication and authorization, which is validated via Picovoice license servers, although the actual voice recognition processing remains 100% offline.

Who it’s for

Developers building cross-platform applications that require private, efficient, and accurate speech-to-text capabilities across desktop (Windows, macOS, Linux), mobile (iOS, Android), web (Chrome, Safari, Firefox, Edge), and embedded devices (Raspberry Pi).

Highlights

  • Privacy-First: All voice processing runs locally on the device.
  • Broad Platform Support: Compatible with a wide range of operating systems, browsers, and hardware.
  • Computationally Efficient: Designed to be compact and efficient in its resource usage.
  • Multi-language Support: Supports English, French, German, Italian, Japanese, Korean, Portuguese, and Spanish.

Related

  • Project
    Picovoice/cheetah

    Cheetah is Picovoice’s on‑device, streaming speech‑to‑text engine. It runs locally on phones, PCs, Raspberry Pi and browsers, supports several major languages, offers word‑level metadata, and provides SDKs for Python, C, iOS, Android, Flutter, React, Node, Java, .NET and Web. An AccessKey unlocks the library; you create an engine, feed PCM audio frames, receive partial transcripts, and flush for the final result. Demos and package builds are included for all platforms.

  • Project
    goodroot/hyprwhspr

    hyprwhspr is a native Linux speech‑to‑text app that runs as a system‑wide service. It uses local models (Whisper, Cohere, Parakeet, Qwen3‑ASR, etc.) for near‑real‑time dictation, optionally falls back to cloud APIs, and integrates with Wayland/X11 desktops, status bars, and a themed visualiser. Install via the AUR or a one‑line script, configure with `hyprwhspr setup`, then toggle dictation with `Super+Alt+D` to have transcribed text auto‑pasted into any window.

  • Project
    matthartman/ghost-pepper

    A private, on-device voice-to-text and meeting transcription app for macOS that uses local AI models to ensure no data leaves the machine.

  • Project
    watzon/pindrop

    Pindrop is a discontinued, native‑macOS menu‑bar dictation app that transcribes speech offline using on‑device models (Whisper‑Kit, Parakeet, SenseVoice, Apple Speech) or optional OpenAI cloud models. It offers real‑time streaming, multilingual support, speaker diarization, a searchable transcript library, AI‑enhanced cleanup, and privacy‑first defaults. Built in Swift/SwiftUI, it runs on macOS 14+ (Apple Silicon recommended) and is replaced by the Sayso app.

  • Project
    OpenWhispr/openwhispr

    A privacy-first, open-source voice-to-text application providing local AI dictation, meeting transcription with speaker diarization, and AI-powered notes.