Best-in-class speech recognition for Node.js on Apple Silicon

Transcribe audio in 99 languages. Run 100% offline on your Mac.

OpenAI's Whisper is the gold standard for speech recognition accuracy. This package brings it to Node.js – powered by Apple's Neural Engine for fast, private, local transcription.

🎯 Accuracy first. Whisper large-v3-turbo delivers state-of-the-art transcription quality – better than any cloud API, right on your Mac.

🌍 99 languages. From Afrikaans to Zulu. Handles accents, dialects, and background noise.

🔒 100% private. Your audio never leaves your device. No API keys. No cloud. No subscription.

⚡ Fast enough. 14x real-time on M1 Ultra – transcribe 1 hour of audio in under 5 minutes.

Running Whisper without hardware acceleration is painfully slow. Here's how the alternatives compare:

The Neural Engine in every Apple Silicon Mac is a dedicated ML accelerator that usually sits idle. This package puts it to work.

Need even more speed? Our sister project parakeet-coreml trades language coverage for 40x real-time performance.

• 🎯 99 Languages – Full OpenAI Whisper multilingual support
• 🚀 14x real-time – 1 hour of audio in ~4.5 minutes (M1 Ultra)
• 🍎 Neural Engine – Runs on Apple's dedicated ML chip via CoreML
• 🔒 Fully Offline – No internet required after setup
• 📦 Zero Dependencies – No Python, no subprocess, no hassle
• 📝 Timestamps – Segment-level timing for subtitles
• ⬇️ One Command Setup – npx whisper-coreml download

# Install
npm install whisper-coreml

# Download the model (~3GB, one-time)
npx whisper-coreml download

Requirements: macOS 14+ (Sonoma), Apple Silicon (M1/M2/M3/M4), Node.js 20+

Measured on M1 Ultra:

5 min audio  →  22 seconds  →  14x real-time
1 hour audio →  4.5 minutes

Run npx whisper-coreml benchmark to test on your machine.

import { WhisperAsrEngine, getModelPath } from "whisper-coreml"

const engine = new WhisperAsrEngine({
  modelPath: getModelPath()
})

await engine.initialize()

// Transcribe audio (16kHz, mono, Float32Array)
const result = await engine.transcribe(audioSamples, 16000)

console.log(result.text)
// "Hello, this is a test transcription."

console.log(`Language: ${result.language}`)
console.log(`Processed in ${result.durationMs}ms`)

// Segments include timestamps
for (const seg of result.segments) {
  console.log(`[${seg.startMs}ms - ${seg.endMs}ms] ${seg.text}`)
}

engine.cleanup()

Example with ffmpeg:

ffmpeg -i input.mp3 -ar 16000 -ac 1 -f f32le output.pcm

Then load the raw PCM file:

import { readFileSync } from "fs"

const buffer = readFileSync("output.pcm")
const samples = new Float32Array(buffer.buffer, buffer.byteOffset, buffer.length / 4)

# Download the model (~1.5GB)
npx whisper-coreml download

# Check status
npx whisper-coreml status

# Run benchmark (requires cloned repo)
npx whisper-coreml benchmark

# Get model directory path
npx whisper-coreml path

The main class for speech recognition.

new WhisperAsrEngine(options: WhisperAsrOptions)

interface TranscriptionResult {
  text: string // Full transcription
  language: string // Detected language (ISO code)
  durationMs: number // Processing time in milliseconds
  segments: TranscriptionSegment[]
}

interface TranscriptionSegment {
  startMs: number // Segment start in milliseconds
  endMs: number // Segment end in milliseconds
  text: string // Transcription for this segment
  confidence: number // Confidence score (0-1)
}

┌─────────────────────────────────────────────────────────┐
│                    Your Node.js App                     │
├─────────────────────────────────────────────────────────┤
│                  whisper-coreml API                     │  TypeScript
├─────────────────────────────────────────────────────────┤
│                   Native Addon                          │  N-API + C++
│                  (whisper_engine)                       │
├─────────────────────────────────────────────────────────┤
│                    whisper.cpp                          │  C++
├─────────────────────────────────────────────────────────┤
│                      CoreML                             │  Apple Framework
├─────────────────────────────────────────────────────────┤
│                 Apple Neural Engine                     │  Dedicated ML Silicon
└─────────────────────────────────────────────────────────┘

• Maximum accuracy – When other solutions aren't good enough
• Rare languages – 99 languages, far beyond English/European
• Accented speech – Whisper handles accents and dialects well
• Noisy audio – Robust to background noise and music

Contributions are welcome! Please read our Contributing Guide for details.

MIT – see LICENSE for details.

• whisper.cpp by Georgi Gerganov
• OpenAI Whisper by OpenAI

Open Source at Sebastian Software
Copyright © 2026 Sebastian Software GmbH