Timbre-Aware Pitch Estimator

TAPE is a novel pitch estimator (or pitch stream segregator) that improves blind pitch estimation with the help of timbre-specific priors. As a proof-of-concept, we present it with a real-world use case on violin--piano duets, one of the most common scenarios in chamber music.

Usage:

from tape.pitch_estimator import TAPE
import torch
import librosa
device = 'cuda' if torch.cuda.is_available() else 'cpu'

violin_tape = TAPE(instrument='violin',   # supported instruments are the violin, the flute and the (Bb soprano) clarinet
                   window_size=16*1024,   # 1.024 seconds,
                   hop_length=128         # 128/16000= 8 miliseconds
                   ).to(device)

audio, _ = librosa.load('violin-piano.wav', sr=violin_tape.sr)
with torch.no_grad():
  time, frequency, confidence, activation = violin_tape.predict(torch.tensor(audio), 
                                                                viterbi=False, batch_size=128)

For real-world violin--piano examples from YouTube, please visit the colab demo. You can also use your own YouTube links to directly sonify the violin pitch track from any recording.