bedcmmPitch 0.2.1

Pitch detection using periodicity analysis

bedcmmPitch

Research / Evaluation / PoC use only.

Pitch Detection Algorithm

This repository provides a pitch detection algorithm based on a periodicity analysis method.
It is intended for research and Proof of Concept (PoC) use. Robust to spike-like noise and impulsive artifacts.

The implementation includes four main functions:

• calc_Pitch: computes pitch values
• calc_Pitch_with_bayes: computes pitch values with bayesain(for on-line)
• calc_Pitch_with_viterbi: computes pitch values with viterbi(for off-line)
• calc_bedcmm: outputs intermediate periodicity analysis results

Installation

Install from PyPI

pip install bedcmmPitch

Optional: Build Cython extension from source

python setup.py build_ext --inplace

Example

import numpy as np
from bedcmmPitch import calc_Pitch

fs = 44100
t = np.arange(fs) / fs
x = np.sin(2*np.pi*440*t)

pitch, score = calc_Pitch(x)

print(np.nanmean(pitch))

Robustness to Spike Noise (Comparison with pYIN)

We compared this method with pYIN implemented in librosa under spike noise conditions.

Results

Time-series comparison (spike ratio = 1000)

• YIN shows large pitch errors under spike noise
• pYIN is more stable but still degrades
• bedcmm remains stable

Detection rate vs spike noise

Detection rate is defined as the ratio of frames where the pitch error is within ±5% of the ground truth.

Experimental Setup

• Signal: 220 Hz sine wave
• Sampling rate: 44.1 kHz
• Spike noise: random impulses with varying amplitude ratio (see simulation_test.py)
• Window size: 2048
• Hop size: 256

Notes

• pYIN uses default parameters from librosa
• bedcmm uses default parameters unless otherwise noted
• Reproducible scripts are available in simulation_test.py.

Summary

bedcmm shows significantly higher robustness to spike noise compared to pYIN.

About bedcmm

bedcmm is a periodicity analysis method used as the core of this pitch detection algorithm.

For more details, please refer to the original repository: https://github.com/YASUHARA-Wataru/bedcmm

Requirements

• numpy

Cython is only required when building from source.

Functions

calc_Pitch

Input (default):

• data : 1D array (signal data)
• fs (44100) : float (sampling rate)
• window_size (2048) : int (window size)
• hop_size (256) : int (hop size)
• fmin(65) : float(calculation min freq)
• fmax(2000) : float(calculation max freq)
• pp_mode ('positive+negative') : str (preprocessing mode: 'positive', 'negative', 'positive+negative', 'threshold_diff')
• pp_threshold (0) : float (preprocessing threshold, used in 'threshold_diff' mode)
• bedcmm_smooth (3) : int (smoothing size for bedcmm result; 1 means no smoothing)
• pitch_detect_mode ('peak') : str (peak detection mode: 'score', 'static', 'maximum', 'peak')
• pitch_detect_thre (0.85) : float (threshold for peak detection)
• interpolator_mode ('parabolic') : str (peak interpolation mode: 'parabolic', 'centroid', 'gaussian', 'no')

Output:

• Pitch data: 1D array
• Score data: 1D array (bedcmm score)

---

calc_Pitch_with_bayes

Input (default):

• data : 1D array data(signal data)
• fs(44100) : float(sampling rate)
• window_size(2048) : int(window size)
• hop_size(256) : int(hop size)
• fmin(65) : float(calculation min freq)
• fmax(2000) : float(calculation max freq)
• pp_mode('positive+negative'): str(perprocessing mode('positive','negative','positive+negative','threshold_diff'))
• pp_threshold(0) : float(perprocessing threshold(using in 'threshold_diff mode'))
• alpha(0.7) : float(bayesain tracking parameter)
• sigma(0.1) : float(pitch conversion parameter)
• interpolator_mode('parabolic'): str(peak index interpolator mode('parabolic','centroid','gaussian', 'no')) Output:
• Pitch data:1D array data(pitch data)
• score data:1D array data(bedcmm score)
• prob data:1D array data(probability)

---

calc_Pitch_with_viterbi

Input (default):

• data : 1D array data(signal data)
• fs(44100) : float(sampling rate)
• window_size(2048) : int(window size)
• hop_size(256) : int(hop size)
• fmin(65) : float(calculation min freq)
• fmax(2000) : float(calculation max freq)
• pp_mode('positive+negative'): str(perprocessing mode('positive','negative','positive+negative','threshold_diff'))
• pp_threshold(0) : float(perprocessing threshold(using in 'threshold_diff mode'))
• beta(10) : float(Viterbi control parameter)
• sigma(0.1) : float(pitch conversion parameter)
• interpolator_mode('parabolic'): str(peak index interpolator mode('parabolic','centroid','gaussian', 'no')) Output:
• Pitch data:1D array data(pitch data)
• score data:1D array data(bedcmm score)
• prob data:1D array data(probability)

---

calc_bedcmm

Input:

• data : 1D array (signal data)
• fs (44100) : float (sampling rate)
• window_size (2048) : int (window size)
• hop_size (256) : int (hop size)
• fmin(None) : float(calculation min freq)
• fmax(None) : float(calculation max freq)
• pp_mode ('positive') : str (preprocessing mode)
• pp_threshold (0) : float (used in 'threshold_diff' mode)

Output:

• bedcmm data: 2D array (time, bedcmm)
• mean data: 1D array (time)

---

Calculation speed

The benchmark script is speed_test.py. Real-time factor ≈ 1.0 indicates near real-time processing. Values may slightly vary depending on the execution environment.

Cython

preprocessing mode 'positive+negative'(default)(Accurate:Robust to noise and outliers)

=== Benchmark Result ===
 audio_sec  proc_sec  real_time_factor
        10   10.9753            1.0975
        20   23.1892            1.1595
        30   31.2636            1.0421

preprocessing mode 'positive' (Fast:Reduced computation, slightly less robust)

=== Benchmark Result ===
 audio_sec  proc_sec  real_time_factor
        10    4.6133            0.4613
        20   11.0052            0.5503
        30   17.7401            0.5913

Notes

It is recommended to use the default parameters.

Default calculation pitch range is 65[Hz] to 2000[Hz].

You may observe differences depending on parameter settings, so feel free to experiment if needed.

• When pitch_detect_mode = 'score', the threshold is calculated as:
  (mean of input signal) × pitch_detect_threshold

• When pitch_detect_mode = 'static', pitch_detect_threshold is used directly as a threshold.

• When pitch_detect_mode = 'peak', the threshold is:
  (maximum peak value of bedcmm) × pitch_detect_threshold

• When pitch_detect_mode = 'maximum', it may not function properly unless pitch_range is restricted.

• The default pp_mode prioritizes accuracy but is computationally expensive.
  For faster computation, you can use 'positive' or 'negative'.

• When using pp_mode = 'threshold_diff', all values must be non-negative; otherwise, the scoring mechanism may not function correctly.

---

for Development fast Installation

1. Copy the bedcmmPitch folder into your working directory

2. Copy setup.py into your working directory

3. Ensure the following libraries are installed:

   • numpy
   • cython
   • setuptools
   • (optional) matplotlib (for visualization)

   If not installed, use pip or uv to install them.

4. In your working directory, run:

python setup.py build_ext --inplace

5. Run main.py to verify functionality

Note: Step 4 can be skipped; the pure Python implementation will still work.

---

License

This repository is available for research, evaluation, and Proof of Concept (PoC) purposes only.

Permitted Use

• Research use
• Algorithm evaluation and validation
• Proof of Concept (PoC)

PoC usage is allowed without time limitation.

Restrictions

The following uses are prohibited:

• Commercial use (including integration into products or services, or any monetized usage)
• Use in production environments (including integration into operational systems)

Commercial Use

If you wish to use this software for commercial purposes or in a production environment,
a separate commercial license agreement is required.

Please contact via Issues or direct email for inquiries.

Disclaimer

This software is provided "AS IS", without warranty of any kind.
The author shall not be held liable for any damages arising from the use of this software.

---

For detailed terms, please refer to the LICENSE file.

---

Contact

For commercial use, please contact us with a brief description of your use case. fapow.contact[at]gmail.com

Patent Notice

This repository includes technology related to a patented method in Japan.

The patent rights are currently granted in Japan only. However, this does not grant any rights for commercial use of this software or the underlying technology.

For any commercial or production use, please contact the author.