sync: masked waveform correlation, stretch sync, manual start time
- Waveform sync no longer treats uncached waveform regions as silence: the envelope extraction now reports a per-window validity mask and the correlation skips invalid windows on either side, improving accuracy for partially cached clips - Add stretch/speed sync: AudioWaveformSync::EstimateStretchAndOffset searches a playback-rate range plus offset, and a new timeline context action 'Synchronize by Waveform (Adjust Speed)' applies the estimated rate as a clip speed change (with undo) when plain offset alignment is inconclusive - Footage properties dialog gains a Source Start Time field so the value used by source-time sync can be viewed and edited manually instead of relying solely on auto-detected metadata; applied via an undo command, with Footage::ClearSourceStartTime() for removal - Regression tests for masked correlation, stretch estimation, the envelope validity mask, and source-start-time set/clear
This commit is contained in:
@@ -38,6 +38,16 @@ public:
|
||||
bool valid = false;
|
||||
};
|
||||
|
||||
struct StretchOffsetResult {
|
||||
// Playback rate the candidate must be played at to align with the
|
||||
// reference (e.g. 2.0 = candidate runs at half speed and needs to be
|
||||
// sped up 2x)
|
||||
double rate = 1.0;
|
||||
int64_t offset_samples = 0;
|
||||
double confidence = 0.0;
|
||||
bool valid = false;
|
||||
};
|
||||
|
||||
static QVector<double> ExtractRmsEnvelope(const core::SampleBuffer &samples,
|
||||
size_t window_samples);
|
||||
|
||||
@@ -50,6 +60,41 @@ public:
|
||||
const QVector<double> &candidate,
|
||||
size_t window_samples,
|
||||
int64_t max_offset_windows);
|
||||
|
||||
/**
|
||||
* @brief Offset estimation that ignores windows flagged as invalid
|
||||
*
|
||||
* @p reference_valid and @p candidate_valid mark which envelope windows
|
||||
* contain real data (e.g. actually cached waveform regions). Windows
|
||||
* flagged false on either side are excluded from the correlation instead
|
||||
* of being treated as silence, which improves accuracy when parts of the
|
||||
* waveform cache have not been generated yet. Empty masks are treated as
|
||||
* "all windows valid".
|
||||
*/
|
||||
static OffsetResult EstimateEnvelopeOffset(const QVector<double> &reference,
|
||||
const QVector<double> &candidate,
|
||||
const QVector<bool> &reference_valid,
|
||||
const QVector<bool> &candidate_valid,
|
||||
size_t window_samples,
|
||||
int64_t max_offset_windows);
|
||||
|
||||
/**
|
||||
* @brief Estimates a playback-rate change plus offset aligning the
|
||||
* candidate to the reference
|
||||
*
|
||||
* The candidate envelope is resampled at each candidate rate in
|
||||
* [min_rate, max_rate] (step rate_step) and correlated against the
|
||||
* reference. rate > 1 means the candidate runs slower than the reference
|
||||
* and must be sped up. The search is O(rates * lags * overlap), so
|
||||
* callers should bound max_offset_windows to a sensible range.
|
||||
*/
|
||||
static StretchOffsetResult
|
||||
EstimateStretchAndOffset(const QVector<double> &reference,
|
||||
const QVector<double> &candidate,
|
||||
const QVector<bool> &reference_valid,
|
||||
const QVector<bool> &candidate_valid, size_t window_samples,
|
||||
int64_t max_offset_windows, double min_rate,
|
||||
double max_rate, double rate_step);
|
||||
};
|
||||
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user