- Generated by
1.15.0
Monophonic fundamental-frequency estimator, after de Cheveigné & Kawahara 2002 [1], steps 1–5. More...
#include <yin_tracker.hpp>
Public Member Functions | |
| void | prepare (double sample_rate) |
| void | reset () |
| void | discard_history () noexcept |
| Constant-time logical reset for audio-thread fault recovery. | |
| void | set_f0_range (double min_hz, double max_hz) |
| Sets the tracked range. | |
| double | f0_min_hz () const |
| double | f0_max_hz () const |
| bool | process (SampleType x) |
| Feeds one sample. | |
| double | f0_hz () const |
| The current estimate. | |
| double | tau_samples () const |
| The refined period in samples, before the Hz conversion. | |
| bool | voiced () const |
| False when the frame carried no stable period (min d' > kYinThreshold). | |
| double | min_cmnd () const |
| The frame's minimum CMND — the tracker's own confidence reading. | |
| int | latency_samples () const |
| Series law 5: the estimate for a frame needs the whole window, so this is W. | |
| int | window_samples () const |
| int | integration_samples () const |
| int | tau_min () const |
| int | tau_max () const |
| double | cost_mac_per_sample () const |
| Amortized analysis cost, so a caller can see before profiling that this is a one-per-track block rather than a per-voice one. | |
Static Public Member Functions | |
| static constexpr int | hop_samples () |
Static Public Attributes | |
| static constexpr double | kF0MinDefault = 80.0 |
| Lowest tracked fundamental. | |
| static constexpr double | kF0MinLimit = 50.0 |
| static constexpr double | kF0MaxLimit = 2000.0 |
| static constexpr double | kF0MaxDefault = 1000.0 |
| Highest tracked fundamental. | |
| static constexpr double | kYinThreshold = 0.10 |
| The absolute threshold on the cumulative mean normalized difference. | |
| static constexpr int | kHop = 256 |
| Analysis hop — the control cadence. | |
| static constexpr int | kMedianTaps = 3 |
| Median taps for octave-flip rejection. | |
| static constexpr int | kMaxMedianTaps = 5 |
Monophonic fundamental-frequency estimator, after de Cheveigné & Kawahara 2002 [1], steps 1–5.
RT contract: prepare(sample_rate) sizes the analysis buffers and is the only allocating call. process, set_*, and reset allocate nothing, take no locks, and never throw. The per-hop analysis is a burst of work on ONE sample in every kHop — see cost_mac_per_sample for what that costs and why you should not instantiate thirty of these.
Steps 1–5 are the paper's. Step 6 (best-local-estimate) is omitted and replaced by a 3-tap median across hops, which is an original heuristic and cheaper — it kills the single-frame octave flips that are YIN's dominant failure mode at the cost of one extra hop of settling. That substitution is a design choice and is NOT a claim to implement the paper's step 6.
|
inline |
Amortized analysis cost, so a caller can see before profiling that this is a one-per-track block rather than a per-voice one.
|
inlinenoexcept |
Constant-time logical reset for audio-thread fault recovery.
|
inline |
The current estimate.
Held through unvoiced frames rather than dropping to zero, so a caller that ignores voiced() still gets a sane number.
|
inline |
|
inline |
|
inlinestaticconstexpr |
|
inline |
|
inline |
Series law 5: the estimate for a frame needs the whole window, so this is W.
There is no zero-latency variant of a period-length autocorrelation. Report it; do not hide it.
|
inline |
The frame's minimum CMND — the tracker's own confidence reading.
|
inline |
|
inline |
Feeds one sample.
Returns true on the samples where a NEW estimate landed, so a caller can re-run its control chain only on those.
|
inline |
|
inline |
Sets the tracked range.
Takes effect at the next prepare, because the window length — and therefore the reported latency and the buffer sizes — are derived from it. Changing latency mid-stream would be a worse surprise than requiring a re-prepare.
|
inline |
|
inline |
|
inline |
The refined period in samples, before the Hz conversion.
|
inline |
False when the frame carried no stable period (min d' > kYinThreshold).
|
inline |
|
staticconstexpr |
Highest tracked fundamental.
Covers soprano and the guitar's upper register. [design parameter] default 1000 Hz, range 500 .. 2000 Hz.
|
staticconstexpr |
|
staticconstexpr |
Lowest tracked fundamental.
Covers a bass/guitar low E (~82 Hz) and a low male voice. Sets the analysis window, and therefore the latency. [design parameter] default 80 Hz, range 50 .. 200 Hz.
|
staticconstexpr |
|
staticconstexpr |
Analysis hop — the control cadence.
256 samples at 48 kHz is ~187 estimates/s. [design parameter] default 256, range 128 .. 512.
|
staticconstexpr |
|
staticconstexpr |
Median taps for octave-flip rejection.
[design parameter] default 3, range {1 (off), 3, 5}.
|
staticconstexpr |
The absolute threshold on the cumulative mean normalized difference.
Published: [1] §III.D uses 0.10. Range 0.05 .. 0.20 is the paper's own sensitivity discussion, not a free knob.
1.15.0