react-native-pitchy

A real-time pitch detection library for React Native with 6 classical algorithms. Cross-platform shared C++ core for iOS and Android.

Features

• 6 algorithms: ACF2+, YIN, MPM, HPS, AMDF, RAPT
• New Architecture (TurboModules) and Old Architecture support
• Zero dependencies: No ML frameworks, no ONNX — pure C++ signal processing
• YIN-FFT: O(N log N) classical pitch detection — low latency, sub-cent precision
• MPM: McLeod Pitch Method — bounded NSDF with reliable threshold selection
• RAPT: Two-stage pitch tracking — gold standard for speech processing
• Cross-platform: Shared C++ core for iOS and Android

Installation

npm install react-native-pitchy

npx pod-install

Rebuild your app after installing the package.

Permissions

Microphone permissions are required. Use react-native-permissions or similar.

• iOS: Microphone
• Android: RECORD_AUDIO

Usage

import Pitchy from 'react-native-pitchy';

// Initialize with algorithm choice
Pitchy.init({
  algorithm: 'YIN',       // 'ACF2+' | 'YIN' | 'MPM' | 'HPS' | 'AMDF' | 'RAPT'
  bufferSize: 4096,
  minVolume: -60,
});

// Start detection
await Pitchy.start();

// Listen to pitch events
const subscription = Pitchy.addListener(({ pitch, confidence, volume }) => {
  console.log(`Pitch: ${pitch} Hz, Confidence: ${confidence}, Volume: ${volume} dB`);
});

// Stop detection
await Pitchy.stop();
subscription.remove();

Algorithms

| Algorithm | Type | Best For | Latency | Accuracy | |-----------|------|----------|---------|----------| | YIN | Classical (FFT) | Guitar tuning | ~1.3ms | 86.0% RPA | | RAPT | Two-stage NCCF | Speech processing | ~2.5ms | 83.7% RPA | | ACF2+ | Classical | General purpose | ~4.2ms | 81.4% RPA | | MPM | Classical (NSDF) | Instrument tuning | ~1.4ms | 79.1% RPA | | AMDF | Time domain | Low-power devices | ~1.7ms | 74.4% RPA | | HPS | Frequency domain | Harmonic-rich signals | ~0.3ms | 55.8% RPA |

YIN

FFT-accelerated autocorrelation with cumulative mean normalized difference. Best for instrument tuning where low latency and high precision on clean signals matter.

MPM (McLeod Pitch Method)

Uses the Normalized Squared Difference Function (NSDF) bounded to [-1, 1], making threshold selection reliable. Key-maxima peak picking with parabolic interpolation. Excellent for musical instruments — used in many professional tuner apps.

HPS (Harmonic Product Spectrum)

Frequency-domain approach that multiplies downsampled magnitude spectra to amplify the fundamental frequency. Fast and effective for harmonic-rich signals (guitar, bass, brass). Uses Hanning windowing and log-domain parabolic interpolation.

AMDF (Average Magnitude Difference Function)

No-multiply pitch detection — uses absolute differences instead of multiplications, making it extremely fast on low-power hardware. Parabolic interpolation for sub-bin precision, with confidence derived from the depth of the AMDF minimum. Best for scenarios where CPU budget is tight.

RAPT (Robust Algorithm for Pitch Tracking)

Two-stage pitch detection: first a coarse search on 4x downsampled signal for fast candidate generation, then refined search on the original signal near coarse candidates. Uses Normalized Cross-Correlation Function (NCCF) at both stages. Gold standard for speech processing — robust voiced/unvoiced detection.

Benchmarks

Measured on Apple Silicon, 4096 samples @ 44.1kHz, 1000 iterations:

| Algorithm | ms/call | FPS | Clean RPA | Noisy RPA | Notes | |-----------|---------|-----|-----------|-----------|-------| | HPS | 0.31 | 3198 | 100%* | 42.1% | Fastest — frequency domain | | YIN | 1.34 | 744 | 100% | 100% | FFT-accelerated CMND | | MPM | 1.35 | 743 | 100% | 100% | FFT-based NSDF | | AMDF | 1.69 | 591 | 100% | 100% | Time-domain, no FFT | | RAPT | 2.48 | 403 | 100% | 100% | Two-stage NCCF | | ACF2+ | 4.24 | 236 | 100% | 89.5% | Legacy time-domain |

* HPS requires harmonic-rich signals (100% on instrument-like signals, 0% on pure sine).

All algorithms run well within real-time requirements (a 4096-sample buffer at 44.1kHz represents ~93ms of audio).

Real-World Audio Accuracy

Tested on 43 real recordings: 32 orchestral instruments from Philharmonia Orchestra (CC BY-SA 3.0) + 11 male singing voice from Human Voice Dataset (MIT):

| Algorithm | Cello | Guitar | Flute | Clarinet | Fr. Horn | Voice | Overall RPA | |-----------|-------|--------|-------|----------|----------|-------|-------------| | YIN | 100% | 100% | 100% | 100% | 100% | 73% | 86.0% | | RAPT | 100% | 100% | 100% | 100% | 100% | 73% | 83.7% | | ACF2+ | 100% | 83% | 100% | 100% | 100% | 64% | 81.4% | | MPM | 100% | 83% | 100% | 100% | 100% | 64% | 79.1% | | AMDF | 100% | 67% | 100% | 100% | 100% | 73% | 74.4% | | HPS | 40% | 83% | 100% | 50% | 50% | 45% | 55.8% |

RPA = Raw Pitch Accuracy (within 50 cents of ground truth). Voice failures are octave errors on higher male voice notes (A3+) where the second harmonic dominates the fundamental — a known limitation of classical pitch detectors.

Need more accuracy?

react-native-pitchy-pro adds 6 advanced algorithms including ML-powered pitch detection:

| Algorithm | Type | ms/call | Accuracy | Best For | |-----------|------|---------|----------|----------| | CREPE | ML (CNN) | — | ~97.8% RPA* | Highest accuracy, state-of-the-art | | PESTO | ML (Lightweight) | — | ~96.1% RPA* | Fast ML, only 29K params | | SwiftF0 | ML (CNN) | — | ~93% RPA* | Noise-robust environments | | pYIN | HMM + Viterbi | 1.27ms | 100% RPA | Vocal melody extraction | | Salience | Harmonic Analysis | 0.39ms | 100% RPA | Instrument tuning with tracking | | SWIPE' | ERB Matching | 0.11ms | 100% RPA | Noise-robust classical |

* ML accuracy from published papers (Kim et al. 2018, Riou et al. 2023). ML latency depends on device and ONNX Runtime.

Pro features:

• 3 ML models via ONNX Runtime (CREPE, PESTO, SwiftF0)
• pYIN with HMM temporal smoothing for stable vocal tracking
• Salience with pitch tracking, octave correction, and ERB weighting
• SWIPE' with sawtooth harmonic template matching

Learn more | Contact for licensing

API

Types

• PitchyAlgorithm: 'ACF2+' | 'YIN' | 'MPM' | 'HPS' | 'AMDF' | 'RAPT'
• PitchyConfig:
  • bufferSize?: Audio buffer size (default: 4096)
  • minVolume?: Minimum volume threshold in dB (default: -60)
  • algorithm?: Detection algorithm (default: 'ACF2+')
• PitchyEvent:
  • pitch: Detected pitch in Hz (-1 if no pitch)
  • confidence: Detection confidence (0.0 - 1.0)
  • volume: Input volume in dB
• PitchyEventCallback: (event: PitchyEvent) => void

Methods

• Pitchy.init(config?) — Initialize with optional configuration
• Pitchy.start() — Start pitch detection (Promise)
• Pitchy.stop() — Stop pitch detection (Promise)
• Pitchy.isRecording() — Check if running (Promise<boolean>)
• Pitchy.addListener(callback) — Listen to onPitchDetected events

Examples

See the example app for a working debug implementation with volume metering, raw data display, and algorithm switching.

Contributing

See the contributing guide.

License

MIT — Copyright (c) 2024 App Vision