Autocorrelation-based C++ pitch detection algorithms with O(nlogn) or lower running time:
- McLeod pitch method - 2005 paper - visualization
- YIN(-FFT) - 2002 paper - visualization
- Probabilistic YIN - 2014 paper
- Probabilistic MPM - my own invention
The size of the FFT used is the same as the size of the input waveform, such that the output is a single pitch for the entire waveform.
Librosa (among other libraries) uses the STFT to create frames of the input waveform, and applies pitch tracking to each frame with a fixed FFT size (typically 2048 or some other power of two). If you want to track the temporal evolution of pitches in sub-sections of the waveform, you have to handle the waveform splitting yourself (look at wav_analyzer for more details).
Dec 27, 2023 🎅 release:
- Removed SWIPE' algorithm
- It is not based on autocorrelation, I skipped it in all of the tests, and my implementation was basically copy-pasted from kylebgorman/swipe: just use their code instead!
- Fix autocorrelation (in YIN and MPM) for power-of-two sizes in FFTS (see ffts issue #65) by using r2c/c2r transforms (addresses bug #72 reported by jeychenne)
- Fix PYIN bugs to pass all test cases (addresses jansommer's comments in pull-request #84)
- Added many more unit tests, all passing (228/228)
- Go: Go implementation of YIN in this repo (for tutorial purposes)
- Rust: Rust implementation of MPM in this repo (for tutorial purposes)
- Python: transcribe is a Python version of MPM for a proof-of-concept of primitive pitch transcription
- Javascript (WebAssembly): pitchlite has WASM modules of MPM/YIN running at realtime speeds in the browser, and also introduces sub-chunk detection to return the overall pitch of the chunk and the temporal sub-sequence of pitches within the chunk
Suggested usage of this library can be seen in the utility wav_analyzer which divides a wav file into chunks of 0.01s and checks the pitch of each chunk. Sample output of wav_analyzer:
std::vector<float> chunk; // chunk of audio
float pitch_mpm = pitch::mpm(chunk, sample_rate);
float pitch_yin = pitch::yin(chunk, sample_rate);
There are unit tests that use sinewaves (both generated with std::sin and with librosa.tone), and instrument tests using txt files containing waveform samples from the University of Iowa MIS recordings:
$ ./build/pitch_tests
Running main() from ./googletest/src/gtest_main.cc
[==========] Running 228 tests from 22 test suites.
[----------] Global test environment set-up.
[----------] 2 tests from MpmSinewaveTestManualAllocFloat
[ RUN ] MpmSinewaveTestManualAllocFloat.OneAllocMultipleFreqFromFile
[ OK ] MpmSinewaveTestManualAllocFloat.OneAllocMultipleFreqFromFile (38 ms)
...
[----------] 5 tests from YinInstrumentTestFloat
...
[ RUN ] YinInstrumentTestFloat.Acoustic_E2_44100
[ OK ] YinInstrumentTestFloat.Acoustic_E2_44100 (1 ms)
[ RUN ] YinInstrumentTestFloat.Classical_FSharp4_48000
[ OK ] YinInstrumentTestFloat.Classical_FSharp4_48000 (58 ms)
[----------] 5 tests from YinInstrumentTestFloat (174 ms total)
...
[----------] 5 tests from MpmInstrumentTestFloat
[ RUN ] MpmInstrumentTestFloat.Violin_A4_44100
[ OK ] MpmInstrumentTestFloat.Violin_A4_44100 (61 ms)
[ RUN ] MpmInstrumentTestFloat.Piano_B4_44100
[ OK ] MpmInstrumentTestFloat.Piano_B4_44100 (24 ms)
...
[==========] 228 tests from 22 test suites ran. (2095 ms total)
[ PASSED ] 228 tests.
All testing files are here - the progressive degradations are described by the respective numbered JSON file, generated using audio-degradation-toolbox. The original clip is a Viola playing E3 from the University of Iowa MIS. The results come from parsing the output of wav_analyzer to count how many 0.1s slices of the input clip were in the ballpark of the expected value of 164.81 - I considered anything 160-169 to be acceptable:
| Degradation level | MPM # correct | YIN # correct |
|---|---|---|
| 0 | 26 | 22 |
| 1 | 23 | 21 |
| 2 | 19 | 21 |
| 3 | 18 | 19 |
| 4 | 19 | 19 |
| 5 | 18 | 19 |
You need Linux, cmake, and gcc (I don't officially support other platforms). The library depends on ffts and mlpack. The tests depend on libnyquist, googletest, and google benchmark. Dependency graph:
Build and install with cmake:
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build "build"
# install to your system
cd build && make install
# run tests and benches
./build/pitch_tests
./build/pitch_bench
# run wav_analyzer
./build/wav_analyzerTo simplify the setup, there's a Dockerfile that sets up a Ubuntu container with all the dependencies for compiling the library and running the included tests and benchmarks:
# build
$ docker build --rm --pull -f "Dockerfile" -t pitchdetection:latest "."
$ docker run --rm --init -it pitchdetection:latestn.b. You can pull the esimkowitz/pitchdetection image from DockerHub, but I can't promise that it's up-to-date.
Read the header and the example wav_analyzer program.
The namespaces are pitch and pitch_alloc. The functions and classes are templated for <double> and <float> support.
The pitch namespace functions perform automatic buffer allocation, while pitch_alloc::{Yin, Mpm} give you a reusable object (useful for computing pitch for multiple uniformly-sized buffers):
#include <pitch_detection.h>
std::vector<double> audio_buffer(8192);
double pitch_yin = pitch::yin<double>(audio_buffer, 48000);
double pitch_mpm = pitch::mpm<double>(audio_buffer, 48000);
double pitch_pyin = pitch::pyin<double>(audio_buffer, 48000);
double pitch_pmpm = pitch::pmpm<double>(audio_buffer, 48000);
pitch_alloc::Mpm<double> ma(8192);
pitch_alloc::Yin<double> ya(8192);
for (int i = 0; i < 10000; ++i) {
auto pitch_yin = ya.pitch(audio_buffer, 48000);
auto pitch_mpm = ma.pitch(audio_buffer, 48000);
auto pitch_pyin = ya.probabilistic_pitch(audio_buffer, 48000);
auto pitch_pmpm = ma.probabilistic_pitch(audio_buffer, 48000);
}