0;jildZddlZddlZddlmZddlmZddlmZddlm Z ddl m Z dd l m Z mZmZmZmZmZmZmZdd lmZdd lmZdd lmZmZmZmZgd Zddddddddddd dejdeeeeefeefde de!deeeeefeefde"dee"dee"ded e!d!ed"eejejffd#Z#ddddddddddd dejdeeeeefeefde de!deeeeefeefde"dee"dee"ded e!d!ed"ejfd$Z$ddddddddddd dejdeeeeefeefde de!deeeeefeefde"dee"dee"ded e!d!ed"ejfd%Z%dejd&e d'e d"ejfd(Z&d)d*dd+dejd,e d-e d.e"d/e'd0e!d'e d"ejfd1Z(dd2dejd3eee"e"fd4e!d"ejfd5Z)dd6d7ej*ej*fdejd8e"de"d9e d:eee fd;ed"ejfd<Z+d7ej*dd6ej*d=dejd9e d:ee efd8e"de"d;ed"eejejffd>Z,d7ej*dd6ej*d=dejd9e d:ee efd8e"de"d;ed"ejfd?Z-ed@d@d@dAdejdBe dCeedDedd"ejf dEZ.ed@d@dFdejdBe dCeedDedd"eejejff dGZ.ed@d@dFdejdBe dCeedDe!d"eejeejejfff dHZ.dIdddAdejdBe dCeedDe!d"eejeejejfff dJZ.ed@d@d@dAdejdBe dCeedDedd"ejf dKZ/ed@d@dFdejdBe dCeedDedd"eejejff dLZ/dIdddAdejdBe dCeedDe!d"eejeejejfff dMZ/dS)Na Effects ======= Harmonic-percussive source separation ------------------------------------- .. autosummary:: :toctree: generated/ hpss harmonic percussive Time and frequency ------------------ .. autosummary:: :toctree: generated/ time_stretch pitch_shift Miscellaneous ------------- .. autosummary:: :toctree: generated/ remix trim split preemphasis deemphasis N)core) decompose)feature)util)ParameterError)AnyCallableIterableOptionalTupleListUnionoverload)Literal) ArrayLike) _WindowSpec _PadModeSTFT _IntLike_co _FloatLike_co)hpssharmonic percussive time_stretch pitch_shiftremixtrimsplit@F?ihannTconstant) kernel_sizepowermaskmarginn_fft hop_length win_lengthwindowcenterpad_modeyr$r%r&r'r(r)r*r+r,r-returnc "tj||||| | } tj| ||||\} } tj| |j|||| |jd}tj| |j|||| |jd}||fS)aDecompose an audio time series into harmonic and percussive components. This function automates the STFT->HPSS->ISTFT pipeline, and ensures that the output waveforms have equal length to the input waveform ``y``. Parameters ---------- y : np.ndarray [shape=(..., n)] audio time series. Multi-channel is supported. kernel_size power mask margin See `librosa.decompose.hpss` n_fft hop_length win_length window center pad_mode See `librosa.stft` Returns ------- y_harmonic : np.ndarray [shape=(..., n)] audio time series of the harmonic elements y_percussive : np.ndarray [shape=(..., n)] audio time series of the percussive elements See Also -------- harmonic : Extract only the harmonic component percussive : Extract only the percussive component librosa.decompose.hpss : HPSS on spectrograms Examples -------- >>> # Extract harmonic and percussive components >>> y, sr = librosa.load(librosa.ex('choice')) >>> y_harmonic, y_percussive = librosa.effects.hpss(y) >>> # Get a more isolated percussive component by widening its margin >>> y_harmonic, y_percussive = librosa.effects.hpss(y, margin=(1.0,5.0)) r(r)r*r,r-r$r%r&r'dtyper(r)r*r,lengthrstftrristftr5shape)r.r$r%r&r'r(r)r*r+r,r-r8 stft_harm stft_percy_harmy_percs I/root/voice-cloning/.venv/lib/python3.11/site-packages/librosa/effects.pyrrBs~ 9     D%> +UfIy Zgwr{FZgwr{F 6>c tj||||| | } tj| ||||d} tj| |j|||| |jd} | S)aExtract harmonic elements from an audio time-series. Parameters ---------- y : np.ndarray [shape=(..., n)] audio time series. Multi-channel is supported. kernel_size power mask margin See `librosa.decompose.hpss` n_fft hop_length win_length window center pad_mode See `librosa.stft` Returns ------- y_harmonic : np.ndarray [shape=(..., n)] audio time series of just the harmonic portion See Also -------- hpss : Separate harmonic and percussive components percussive : Extract only the percussive component librosa.decompose.hpss : HPSS for spectrograms Examples -------- >>> # Extract harmonic component >>> y, sr = librosa.load(librosa.ex('choice')) >>> y_harmonic = librosa.effects.harmonic(y) >>> # Use a margin > 1.0 for greater harmonic separation >>> y_harmonic = librosa.effects.harmonic(y, margin=3.0) r1r2rr3r4r7)r.r$r%r&r'r(r)r*r+r,r-r8r;r=s r?rrt 9     D +Uf I Zgwr{F Mr@c tj||||| | } tj| ||||d} tj| |j|||| |jd} | S)aExtract percussive elements from an audio time-series. Parameters ---------- y : np.ndarray [shape=(..., n)] audio time series. Multi-channel is supported. kernel_size power mask margin See `librosa.decompose.hpss` n_fft hop_length win_length window center pad_mode See `librosa.stft` Returns ------- y_percussive : np.ndarray [shape=(..., n)] audio time series of just the percussive portion See Also -------- hpss : Separate harmonic and percussive components harmonic : Extract only the harmonic component librosa.decompose.hpss : HPSS for spectrograms Examples -------- >>> # Extract percussive component >>> y, sr = librosa.load(librosa.ex('choice')) >>> y_percussive = librosa.effects.percussive(y) >>> # Use a margin > 1.0 for greater percussive separation >>> y_percussive = librosa.effects.percussive(y, margin=3.0) r1r2rr3r4r7)r.r$r%r&r'r(r)r*r+r,r-r8r<r>s r?rrrBr@ratekwargsc \|dkrtdtj|fi|}tj|||dd|dd}t t |jd|z }tj|f|j |d|}|S) aTime-stretch an audio series by a fixed rate. Parameters ---------- y : np.ndarray [shape=(..., n)] audio time series. Multi-channel is supported. rate : float > 0 [scalar] Stretch factor. If ``rate > 1``, then the signal is sped up. If ``rate < 1``, then the signal is slowed down. **kwargs : additional keyword arguments. See `librosa.decompose.stft` for details. Returns ------- y_stretch : np.ndarray [shape=(..., round(n/rate))] audio time series stretched by the specified rate See Also -------- pitch_shift : pitch shifting librosa.phase_vocoder : spectrogram phase vocoder pyrubberband.pyrb.time_stretch : high-quality time stretching using RubberBand Examples -------- Compress to be twice as fast >>> y, sr = librosa.load(librosa.ex('choice')) >>> y_fast = librosa.effects.time_stretch(y, rate=2.0) Or half the original speed >>> y_slow = librosa.effects.time_stretch(y, rate=0.5) rzrate must be a positive numberr)Nr()rDr)r(r3)r5r6) rrr8 phase_vocodergetintroundr:r9r5)r.rDrEr8 stft_stretch len_stretch y_stretchs r?rrRsL qyy=>>> 9Q ! !& ! !D% ::lD11jj$'' LeAGBK$.//00K <Uqw{UUfUUI r@ soxr_hq)bins_per_octaveres_typescalesrn_stepsrPrQrRc *tj|std|ddt| |z z}t jt |fd|i|t||z |||}tj||jdS)aShift the pitch of a waveform by ``n_steps`` steps. A step is equal to a semitone if ``bins_per_octave`` is set to 12. Parameters ---------- y : np.ndarray [shape=(..., n)] audio time series. Multi-channel is supported. sr : number > 0 [scalar] audio sampling rate of ``y`` n_steps : float [scalar] how many (fractional) steps to shift ``y`` bins_per_octave : int > 0 [scalar] how many steps per octave res_type : string Resample type. By default, 'soxr_hq' is used. See `librosa.resample` for more information. scale : bool Scale the resampled signal so that ``y`` and ``y_hat`` have approximately equal total energy. **kwargs : additional keyword arguments. See `librosa.decompose.stft` for details. Returns ------- y_shift : np.ndarray [shape=(..., n)] The pitch-shifted audio time-series See Also -------- time_stretch : time stretching librosa.phase_vocoder : spectrogram phase vocoder pyrubberband.pyrb.pitch_shift : high-quality pitch shifting using RubberBand Examples -------- Shift up by a major third (four steps if ``bins_per_octave`` is 12) >>> y, sr = librosa.load(librosa.ex('choice')) >>> y_third = librosa.effects.pitch_shift(y, sr=sr, n_steps=4) Shift down by a tritone (six steps if ``bins_per_octave`` is 12) >>> y_tritone = librosa.effects.pitch_shift(y, sr=sr, n_steps=-6) Shift up by 3 quarter-tones >>> y_three_qt = librosa.effects.pitch_shift(y, sr=sr, n_steps=3, ... bins_per_octave=24) zbins_per_octave=z must be a positive integer.r rD)orig_sr target_srrQrRr3)size) ris_positive_intrfloatrresampler fix_lengthr:) r.rSrTrPrQrRrErDy_shifts r?rrsL   0 0  L L L L    E'NN?_4 5DmQ,,T,V,,b D  G ?7 5 5 55r@) align_zeros intervalsr^cg}|rctj|}tjtj|d}tj|t |g}|D]J}|r|tj||}||d|d|dfKtj |dS)a^Remix an audio signal by re-ordering time intervals. Parameters ---------- y : np.ndarray [shape=(..., t)] Audio time series. Multi-channel is supported. intervals : iterable of tuples (start, end) An iterable (list-like or generator) where the ``i``th item ``intervals[i]`` indicates the start and end (in samples) of a slice of ``y``. align_zeros : boolean If ``True``, interval boundaries are mapped to the closest zero-crossing in ``y``. If ``y`` is stereo, zero-crossings are computed after converting to mono. Returns ------- y_remix : np.ndarray [shape=(..., d)] ``y`` remixed in the order specified by ``intervals`` Examples -------- Load in the example track and reverse the beats >>> y, sr = librosa.load(librosa.ex('choice')) Compute beats >>> _, beat_frames = librosa.beat.beat_track(y=y, sr=sr, ... hop_length=512) Convert from frames to sample indices >>> beat_samples = librosa.frames_to_samples(beat_frames) Generate intervals from consecutive events >>> intervals = librosa.util.frame(beat_samples, frame_length=2, ... hop_length=1).T Reverse the beat intervals >>> y_out = librosa.effects.remix(y, intervals[::-1]) r3.rraxis) rto_mononpnonzerozero_crossingsappendlenr match_events concatenate)r.r_r^y_outy_monozerosintervals r?rrs^ E0a 4.v6677; %#f++//88  AT.x??@H QsHQK(1+5567777 >%b ) ) ))r@i< frame_lengthtop_dbref aggregatec ^tj|||}tj|ddddf|d}|jdkrct j||t|jdz }t j|tt|jdz }|| kS)aFrame-wise non-silent indicator for audio input. This is a helper function for `trim` and `split`. Parameters ---------- y : np.ndarray Audio signal, mono or stereo frame_length : int > 0 The number of samples per frame hop_length : int > 0 The number of samples between frames top_db : number The threshold (in decibels) below reference to consider as silence. You can also use a negative value for `top_db` to treat any value below `ref + |top_db|` as silent. This will only make sense if `ref` is not `np.max`. ref : callable or float The reference amplitude aggregate : callable [default: np.max] Function to aggregate dB measurements across channels (if y.ndim > 1) Note: for multiple leading axes, this is performed using ``np.apply_over_axes``. Returns ------- non_silent : np.ndarray, shape=(m,), dtype=bool Indicator of non-silent frames )r.rpr).rN)rrrqrra) rrmsramplitude_to_dbndimrdapply_over_axesrangesqueezetuple)r.rpr)rqrrrsmsedbs r?_signal_to_frame_nonsilentr~)sX +  L L LC)#c1aaai.c$OOOB w{{   2uRWq[/A/A B BZuRWq['9'9!:!: ; ; ; <r@)rqrrrpr)rsc t||||||}tj|}|jdkrot t j|d|}t|jdt t j|ddz|} nd\}} |d|| ftj || gfS)a<Trim leading and trailing silence from an audio signal. Silence is defined as segments of the audio signal that are `top_db` decibels (or more) quieter than a reference level, `ref`. By default, `ref` is set to the signal's maximum RMS value. It's important to note that if the entire signal maintains a uniform RMS value, there will be no segments considered quieter than the maximum, leading to no trimming. This implies that a completely silent signal will remain untrimmed with the default `ref` setting. In these situations, an explicit value for `ref` (in decibels) should be used instead. Parameters ---------- y : np.ndarray, shape=(..., n) Audio signal. Multi-channel is supported. top_db : number The threshold (in decibels) below reference to consider as silence. You can also use a negative value for `top_db` to treat any value below `ref + |top_db|` as silent. This will only make sense if `ref` is not `np.max`. ref : number or callable The reference amplitude. By default, it uses `np.max` and compares to the peak amplitude in the signal. frame_length : int > 0 The number of samples per analysis frame hop_length : int > 0 The number of samples between analysis frames aggregate : callable [default: np.max] Function to aggregate across channels (if y.ndim > 1) Returns ------- y_trimmed : np.ndarray, shape=(..., m) The trimmed signal index : np.ndarray, shape=(2,) the interval of ``y`` corresponding to the non-silent region: ``y_trimmed = y[index[0]:index[1]]`` (for mono) or ``y_trimmed = y[:, index[0]:index[1]]`` (for stereo). Examples -------- >>> # Load some audio >>> y, sr = librosa.load(librosa.ex('choice')) >>> # Trim the beginning and ending silence >>> yt, index = librosa.effects.trim(y) >>> # Print the durations >>> print(librosa.get_duration(y, sr=sr), librosa.get_duration(yt, sr=sr)) 25.025986394557822 25.007891156462584 rpr)rrrqrsrr)r3r)rr.) r~rd flatnonzerorXrIrframes_to_samplesminr:asarray) r.rqrrrpr)rs non_silentrestartends r?rrdsv, !  JnZ((G|aD*71:*MMMNN GBK &wr{Q:NNN O O    s S%)^ bj%66 66r@cNt||||||}tjtj|t }|dzg}|dr)|dtjdg|dr5|tjt|gtj tj ||}tj ||jd}|d}|S)anSplit an audio signal into non-silent intervals. Parameters ---------- y : np.ndarray, shape=(..., n) An audio signal. Multi-channel is supported. top_db : number > 0 The threshold (in decibels) below reference to consider as silence ref : number or callable The reference amplitude. By default, it uses `np.max` and compares to the peak amplitude in the signal. frame_length : int > 0 The number of samples per analysis frame hop_length : int > 0 The number of samples between analysis frames aggregate : callable [default: np.max] Function to aggregate across channels (if y.ndim > 1) Returns ------- intervals : np.ndarray, shape=(m, 2) ``intervals[i] == (start_i, end_i)`` are the start and end time (in samples) of non-silent interval ``i``. rrrr3r)r3)r~rdrdiffastyperIinsertarrayrgrhrrrjminimumr:reshape)r.rqrrrpr)rsredgess r?rrsD, !  J N27:#4#4S#9#9:: ; ;EQYKE!}' Q! &&&"~2 RXs:/00111  "2>%#8#8Z P P PE Juagbk * *E MM' " "E Lr@.)coefzi return_zfrrrcdSNr.rrrs r? preemphasisr r@)rrcdSrrrs r?rr  %(Cr@cdSrrrs r?rrs 8;sr@g ףp= ?c htjd| g|j}tjdg|j}|d|dddfz|dddfz }tj|}tj|||tj||j\}}|r||fS|S) a Pre-emphasize an audio signal with a first-order differencing filter: y[n] -> y[n] - coef * y[n-1] Parameters ---------- y : np.ndarray [shape=(..., n)] Audio signal. Multi-channel is supported. coef : positive number Pre-emphasis coefficient. Typical values of ``coef`` are between 0 and 1. At the limit ``coef=0``, the signal is unchanged. At ``coef=1``, the result is the first-order difference of the signal. The default (0.97) matches the pre-emphasis filter used in the HTK implementation of MFCCs [#]_. .. [#] https://htk.eng.cam.ac.uk/ zi : number Initial filter state. When making successive calls to non-overlapping frames, this can be set to the ``zf`` returned from the previous call. (See example below.) By default ``zi`` is initialized as ``2*y[0] - y[1]``. return_zf : boolean If ``True``, return the final filter state. If ``False``, only return the pre-emphasized signal. Returns ------- y_out : np.ndarray pre-emphasized signal zf : number if ``return_zf=True``, the final filter state is also returned Examples -------- Apply a standard pre-emphasis filter >>> import matplotlib.pyplot as plt >>> y, sr = librosa.load(librosa.ex('trumpet')) >>> y_filt = librosa.effects.preemphasis(y) >>> # and plot the results for comparison >>> S_orig = librosa.amplitude_to_db(np.abs(librosa.stft(y)), ref=np.max, top_db=None) >>> S_preemph = librosa.amplitude_to_db(np.abs(librosa.stft(y_filt)), ref=np.max, top_db=None) >>> fig, ax = plt.subplots(nrows=2, sharex=True, sharey=True) >>> librosa.display.specshow(S_orig, y_axis='log', x_axis='time', ax=ax[0]) >>> ax[0].set(title='Original signal') >>> ax[0].label_outer() >>> img = librosa.display.specshow(S_preemph, y_axis='log', x_axis='time', ax=ax[1]) >>> ax[1].set(title='Pre-emphasized signal') >>> fig.colorbar(img, ax=ax, format="%+2.f dB") Apply pre-emphasis in pieces for block streaming. Note that the second block initializes ``zi`` with the final state ``zf`` returned by the first call. >>> y_filt_1, zf = librosa.effects.preemphasis(y[:1000], return_zf=True) >>> y_filt_2, zf = librosa.effects.preemphasis(y[1000:], zi=zf, return_zf=True) >>> np.allclose(y_filt, np.concatenate([y_filt_1, y_filt_2])) True See Also -------- deemphasis r!r5Nr.rrr)rdrr5 atleast_1dscipysignallfilter)r.rrrbarkz_fs r?rrsX C$>> x[i] = y[i] + coef * x[i-1] >>> x = deemphasis(y, coef=coef, zi=zi) Parameters ---------- y : np.ndarray [shape=(..., n)] Audio signal. Multi-channel is supported. coef : positive number Pre-emphasis coefficient. Typical values of ``coef`` are between 0 and 1. At the limit ``coef=0``, the signal is unchanged. At ``coef=1``, the result is the first-order difference of the signal. The default (0.97) matches the pre-emphasis filter used in the HTK implementation of MFCCs [#]_. .. [#] https://htk.eng.cam.ac.uk/ zi : number Initial filter state. If inverting a previous preemphasis(), the same value should be used. By default ``zi`` is initialized as ``((2 - coef) * y[0] - y[1]) / (3 - coef)``. This value corresponds to the transformation of the default initialization of ``zi`` in ``preemphasis()``, ``2*x[0] - x[1]``. return_zf : boolean If ``True``, return the final filter state. If ``False``, only return the pre-emphasized signal. Returns ------- y_out : np.ndarray de-emphasized signal zf : number if ``return_zf=True``, the final filter state is also returned Examples -------- Apply a standard pre-emphasis filter and invert it with de-emphasis >>> y, sr = librosa.load(librosa.ex('trumpet')) >>> y_filt = librosa.effects.preemphasis(y) >>> y_deemph = librosa.effects.deemphasis(y_filt) >>> np.allclose(y, y_deemph) True See Also -------- preemphasis r!rNr3rrr.r) rdrr5rmlistr:rrrarangerr)r.rrrrrrkzfs r?rrsI@ #uQW---A #ag&&&A z Xd173B3<((A3.ag > > >L((AqR(88 r $h!C1H+ %#qs( 34x ry--- / ]2  L((AqRYYqw5G5G(HH rby r@)0__doc__numpyrd scipy.signalrrrrrutil.exceptionsrtypingr r r r r rrrtyping_extensionsr numpy.typingr_typingrrrr__all__ndarrayrZboolrIrrrrstrrrmaxr~rrrrrr@r?rs B++++++RRRRRRRRRRRRRRRRRRRR%%%%%%""""""   "   $ $ '!aaa zaU; 34d;6GGa  a a u]M9:D