CREPE Pitch Estimation
CREPE is a deep-learning model that estimates the fundamental frequency (pitch) of a monophonic audio signal directly from its raw waveform.
Overview
It set a new accuracy standard for pitch tracking, especially on noisy or difficult recordings.
Deep Dive
CREPE (Convolutional Representation for Pitch Estimation), introduced in 2018 by Kim, Salamon, Li, and Bello, predicts the pitch of single-note (monophonic) audio such as a sung voice or solo instrument. Unlike classic algorithms like YIN or pYIN that rely on autocorrelation of the signal, CREPE is a deep convolutional neural network trained directly on time-domain audio frames. It frames pitch estimation as a classification problem: it outputs a probability distribution over 360 pitch bins spanning roughly six octaves, each 20 cents apart. The bin with the highest activation, refined with a local weighted average, gives the estimated frequency plus a confidence score. CREPE proved markedly more robust than signal-processing methods, particularly under noise, and is now a standard component in many music and speech analysis pipelines.
Technical Insight
CREPE takes a 1024-sample audio frame and passes it through six stacked convolutional layers, ending in a 360-unit output layer with sigmoid activations. Each unit corresponds to a pitch bin spaced 20 cents apart across about six octaves. The network is trained with binary cross-entropy against a Gaussian-blurred target centered on the true pitch. At inference, the predicted frequency is the local weighted average of activations around the peak bin, and the peak height serves as a confidence value.
Strategic Impact
Access and reach
It improves accessibility through transcription, narration, and voice interfaces.
Cost and budget
Media teams can ship polished audio faster with smaller budgets.
Speed and scale
Customer-facing systems can process spoken interactions at larger scale.
The Future of CREPE Pitch Estimation
Pitch estimation is moving toward joint models that handle polyphony (multiple simultaneous notes), lower latency for real-time tuning and auto-harmony, and smaller distilled networks that run on phones and embedded devices. CREPE's confidence outputs are increasingly fed into downstream tasks like automatic transcription, vocal correction, and expressive performance analysis. Self-supervised and multitask approaches that learn pitch alongside timbre and articulation are likely to extend CREPE-style accuracy beyond clean monophonic audio.
Real-World Implementation
Tracking a singer's pitch for real-time tuning feedback in vocal-training apps
Driving auto-tune and pitch-correction tools with accurate fundamental-frequency curves
Transcribing solo instrument melodies into MIDI or sheet music
Analyzing intonation and vibrato in music education and performance research
Risks & Guardrails
Voice misuse and impersonation risks increase when consent is missing.
Accuracy can drop across accents, dialects, or noisy environments.
Synthetic audio can be mistaken for authentic speech without clear labeling.
Implementation Roadmap
Obtain explicit consent for voice capture, cloning, and reuse.
Test quality across diverse speakers and background conditions.
Define when a human must review or approve outputs.
Label synthetic audio and keep provenance records for accountability.
Keep Exploring
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the CREPE Pitch Estimation quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Next guide
FastPitch Pitch-Controllable TTS
Frequently asked questions
What is CREPE Pitch Estimation?
CREPE is a deep-learning model that estimates the fundamental frequency (pitch) of a monophonic audio signal directly from its raw waveform. It set a new accuracy standard for pitch tracking, especially on noisy or difficult recordings.
What does CREPE estimate from an audio signal?
CREPE predicts the fundamental frequency, i.e. the perceived pitch, of monophonic audio.
CREPE is designed primarily for what kind of audio?
CREPE estimates pitch for monophonic signals such as a single voice or solo instrument.
How does CREPE frame the pitch-estimation task internally?
CREPE outputs a probability distribution over 360 pitch bins, treating estimation as classification.
What does CREPE take as its direct input?
Unlike spectrogram-based methods, CREPE operates on raw waveform frames through convolutional layers.
Approximately how far apart are CREPE's pitch bins?
CREPE uses 360 bins spaced 20 cents apart, giving sub-semitone resolution across about six octaves.