Whisped

A Rust CLI tool for Whisper ASR with automatic model download.

About

Whisped is a command-line interface for OpenAI's Whisper automatic speech recognition (ASR) system. It supports transcribing audio from various file formats, with automatic model downloading and caching for convenience.

Features

• Automatic model downloading and caching
• Support for WAV, MP3, and OGG audio formats
• Multiple model sizes (tiny, base, small, medium, large)
• Optional quantization for faster inference
• Language detection or specification
• Translation to English
• Optional timestamp inclusion

Installation

Prerequisites

• Rust and Cargo (https://rustup.rs/)

Build from source

git clone https://github.com/dragynfruit/whisped.git
cd whisped
cargo build --release

The binary will be available at target/release/whisped.

Usage

# Basic usage
whisped --input audio.mp3 --output transcript.txt

# Specify a model
whisped --model medium --input audio.mp3 --output transcript.txt

# Use quantization for faster inference
whisped --model medium --quant q5_0 --input audio.mp3 --output transcript.txt

# Specify language and include timestamps
whisped --input audio.mp3 --output transcript.txt --language en --timestamps

# Translate to English
whisped --input audio_in_any_language.mp3 --output transcript.txt --translate

Command Line Options

• --model: Model to use (tiny, base, small, medium, large). Default: base
• --quant: Quantization level (e.g., q4_0, q5_0, q8_0). Default: none
• --input: Path to the input audio file (required)
• --output: Path to the output text file (required)
• --language: Language code (e.g., en, fr, de). Default: auto-detect
• --translate: Enable translation to English
• --timestamps: Include timestamps in the output

Acknowledgements

• This tool uses whisper-rs for Rust bindings to OpenAI's Whisper model
• Models are automatically downloaded from HuggingFace