whisperforge-core

GPU-accelerated Whisper model inference with streaming audio, quantization, and KV-cached decoding.

Quick Links

• Full Documentation: WhisperForge Repository
• Installation: See Installation Guide
• Examples: Library Usage

Features

• All Whisper model sizes (tiny.en through large-v2/v3)
• GPU acceleration via WGPU (Vulkan/DX12/Metal)
• burn-flex backend: CPU + automatic GPU dispatch
• INT8 quantization (~4× compression)
• Streaming audio pipeline with resampling
• KV-cache O(n) decoder
• Per-token timestamps via cross-attention

Usage

use whisperforge_core::{Model, WhisperConfig};
use std::path::Path;

let config = WhisperConfig::tiny_en();
let model = Model::load(Path::new("models/tiny_en_converted"))?;
let transcript = model.transcribe(audio_samples, sample_rate)?;
println!("{}", transcript);

See Also

• whisperforge — Command-line binary (wf) including wf convert for HuggingFace model conversion
• whisperforge-align — VAD & SRT output
• whisperforge-diarize — Speaker diarization

For full documentation, visit the WhisperForge repository.