Category N: Speech Recognition

This category covers speech recognition using whisper.apr, a pure Rust implementation of OpenAI's Whisper model.

Recipes

Key Concepts

whisper.apr

A pure Rust implementation of Whisper designed for:

• WASM-first: Runs in browsers without server
• Int4/Int8 Quantization: 4x smaller models
• Streaming: Real-time transcription
• APR v2 Format: Fast loading with LZ4 compression

Model Sizes

Stack Integration

use whisper_apr::{WhisperModel, Transcriber, TranscribeOptions};

// Load quantized model from APR v2
const MODEL: &[u8] = include_bytes!("whisper-small-int8.apr");
let model = WhisperModel::from_apr_bytes(MODEL)?;

// Create transcriber
let transcriber = Transcriber::new(model);

// Transcribe audio
let result = transcriber.transcribe_file("audio.wav", TranscribeOptions::default())?;
println!("Text: {}", result.text);
println!("Language: {} ({:.1}%)", result.language, result.confidence * 100.0);

Falsifiable Claims

Toyota Way Principles

• Jidoka: Stop on unrecognizable audio
• Muda: Quantization eliminates size waste
• Heijunka: Streaming levels processing load