Caution
You're reading an old version of this documentation. If you want up-to-date information, please have a look at 0.10.2.
Source code for librosa.core.audio
#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""Core IO, DSP and utility functions."""
import pathlib
import warnings
import soundfile as sf
import audioread
import numpy as np
import scipy.signal
import resampy
from numba import jit
from .fft import get_fftlib
from .convert import frames_to_samples, time_to_samples
from .._cache import cache
from .. import util
from ..util.exceptions import ParameterError
from ..util.decorators import deprecate_positional_args
__all__ = [
"load",
"stream",
"to_mono",
"resample",
"get_duration",
"get_samplerate",
"autocorrelate",
"lpc",
"zero_crossings",
"clicks",
"tone",
"chirp",
"mu_compress",
"mu_expand",
]
# Resampling bandwidths as percentage of Nyquist
BW_BEST = resampy.filters.get_filter("kaiser_best")[2]
BW_FASTEST = resampy.filters.get_filter("kaiser_fast")[2]
# -- CORE ROUTINES --#
# Load should never be cached, since we cannot verify that the contents of
# 'path' are unchanged across calls.
[docs]@deprecate_positional_args
def load(
path,
*,
sr=22050,
mono=True,
offset=0.0,
duration=None,
dtype=np.float32,
res_type="kaiser_best",
):
"""Load an audio file as a floating point time series.
Audio will be automatically resampled to the given rate
(default ``sr=22050``).
To preserve the native sampling rate of the file, use ``sr=None``.
Parameters
----------
path : string, int, pathlib.Path, soundfile.SoundFile, audioread object, or file-like object
path to the input file.
Any codec supported by `soundfile` or `audioread` will work.
Any string file paths, or any object implementing Python's
file interface (e.g. `pathlib.Path`) are supported as `path`.
If the codec is supported by `soundfile`, then `path` can also be
an open file descriptor (int) or an existing `soundfile.SoundFile` object.
Pre-constructed audioread decoders are also supported here, see the example
below. This can be used, for example, to force a specific decoder rather
than relying upon audioread to select one for you.
sr : number > 0 [scalar]
target sampling rate
'None' uses the native sampling rate
mono : bool
convert signal to mono
offset : float
start reading after this time (in seconds)
duration : float
only load up to this much audio (in seconds)
dtype : numeric type
data type of ``y``
res_type : str
resample type (see note)
.. note::
By default, this uses `resampy`'s high-quality mode ('kaiser_best').
For alternative resampling modes, see `resample`
.. note::
`audioread` may truncate the precision of the audio data to 16 bits.
See :ref:`ioformats` for alternate loading methods.
Returns
-------
y : np.ndarray [shape=(n,) or (..., n)]
audio time series. Multi-channel is supported.
sr : number > 0 [scalar]
sampling rate of ``y``
Examples
--------
>>> # Load an ogg vorbis file
>>> filename = librosa.ex('trumpet')
>>> y, sr = librosa.load(filename)
>>> y
array([-1.407e-03, -4.461e-04, ..., -3.042e-05, 1.277e-05],
dtype=float32)
>>> sr
22050
>>> # Load a file and resample to 11 KHz
>>> filename = librosa.ex('trumpet')
>>> y, sr = librosa.load(filename, sr=11025)
>>> y
array([-8.746e-04, -3.363e-04, ..., -1.301e-05, 0.000e+00],
dtype=float32)
>>> sr
11025
>>> # Load 5 seconds of a file, starting 15 seconds in
>>> filename = librosa.ex('brahms')
>>> y, sr = librosa.load(filename, offset=15.0, duration=5.0)
>>> y
array([0.146, 0.144, ..., 0.128, 0.015], dtype=float32)
>>> sr
22050
>>> # Load using an already open SoundFile object
>>> import soundfile
>>> sfo = soundfile.SoundFile(librosa.ex('brahms'))
>>> y, sr = librosa.load(sfo)
>>> # Load using an already open audioread object
>>> import audioread.ffdec # Use ffmpeg decoder
>>> aro = audioread.ffdec.FFmpegAudioFile(librosa.ex('brahms'))
>>> y, sr = librosa.load(aro)
"""
if isinstance(path, tuple(audioread.available_backends())):
# Force the audioread loader if we have a reader object already
y, sr_native = __audioread_load(path, offset, duration, dtype)
else:
# Otherwise try soundfile first, and then fall back if necessary
try:
y, sr_native = __soundfile_load(path, offset, duration, dtype)
except RuntimeError as exc:
# If soundfile failed, try audioread instead
if isinstance(path, (str, pathlib.PurePath)):
warnings.warn("PySoundFile failed. Trying audioread instead.", stacklevel=2)
y, sr_native = __audioread_load(path, offset, duration, dtype)
else:
raise exc
# Final cleanup for dtype and contiguity
if mono:
y = to_mono(y)
if sr is not None:
y = resample(y, orig_sr=sr_native, target_sr=sr, res_type=res_type)
else:
sr = sr_native
return y, sr
def __soundfile_load(path, offset, duration, dtype):
"""Load an audio buffer using soundfile."""
if isinstance(path, sf.SoundFile):
# If the user passed an existing soundfile object,
# we can use it directly
context = path
else:
# Otherwise, create the soundfile object
context = sf.SoundFile(path)
with context as sf_desc:
sr_native = sf_desc.samplerate
if offset:
# Seek to the start of the target read
sf_desc.seek(int(offset * sr_native))
if duration is not None:
frame_duration = int(duration * sr_native)
else:
frame_duration = -1
# Load the target number of frames, and transpose to match librosa form
y = sf_desc.read(frames=frame_duration, dtype=dtype, always_2d=False).T
return y, sr_native
def __audioread_load(path, offset, duration, dtype):
"""Load an audio buffer using audioread.
This loads one block at a time, and then concatenates the results.
"""
y = []
if isinstance(path, tuple(audioread.available_backends())):
# If we have an audioread object already, don't bother opening
reader = path
else:
# If the input was not an audioread object, try to open it
reader = audioread.audio_open(path)
with reader as input_file:
sr_native = input_file.samplerate
n_channels = input_file.channels
s_start = int(np.round(sr_native * offset)) * n_channels
if duration is None:
s_end = np.inf
else:
s_end = s_start + (int(np.round(sr_native * duration)) * n_channels)
n = 0
for frame in input_file:
frame = util.buf_to_float(frame, dtype=dtype)
n_prev = n
n = n + len(frame)
if n < s_start:
# offset is after the current frame
# keep reading
continue
if s_end < n_prev:
# we're off the end. stop reading
break
if s_end < n:
# the end is in this frame. crop.
frame = frame[: s_end - n_prev]
if n_prev <= s_start <= n:
# beginning is in this frame
frame = frame[(s_start - n_prev) :]
# tack on the current frame
y.append(frame)
if y:
y = np.concatenate(y)
if n_channels > 1:
y = y.reshape((-1, n_channels)).T
else:
y = np.empty(0, dtype=dtype)
return y, sr_native
[docs]@deprecate_positional_args
def stream(
path,
*,
block_length,
frame_length,
hop_length,
mono=True,
offset=0.0,
duration=None,
fill_value=None,
dtype=np.float32,
):
"""Stream audio in fixed-length buffers.
This is primarily useful for processing large files that won't
fit entirely in memory at once.
Instead of loading the entire audio signal into memory (as
in `load`, this function produces *blocks* of audio spanning
a fixed number of frames at a specified frame length and hop
length.
While this function strives for similar behavior to `load`,
there are a few caveats that users should be aware of:
1. This function does not return audio buffers directly.
It returns a generator, which you can iterate over
to produce blocks of audio. A *block*, in this context,
refers to a buffer of audio which spans a given number of
(potentially overlapping) frames.
2. Automatic sample-rate conversion is not supported.
Audio will be streamed in its native sample rate,
so no default values are provided for ``frame_length``
and ``hop_length``. It is recommended that you first
get the sampling rate for the file in question, using
`get_samplerate`, and set these parameters accordingly.
3. Many analyses require access to the entire signal
to behave correctly, such as `resample`, `cqt`, or
`beat_track`, so these methods will not be appropriate
for streamed data.
4. The ``block_length`` parameter specifies how many frames
of audio will be produced per block. Larger values will
consume more memory, but will be more efficient to process
down-stream. The best value will ultimately depend on your
application and other system constraints.
5. By default, most librosa analyses (e.g., short-time Fourier
transform) assume centered frames, which requires padding the
signal at the beginning and end. This will not work correctly
when the signal is carved into blocks, because it would introduce
padding in the middle of the signal. To disable this feature,
use ``center=False`` in all frame-based analyses.
See the examples below for proper usage of this function.
Parameters
----------
path : string, int, sf.SoundFile, or file-like object
path to the input file to stream.
Any codec supported by `soundfile` is permitted here.
An existing `soundfile.SoundFile` object may also be provided.
block_length : int > 0
The number of frames to include in each block.
Note that at the end of the file, there may not be enough
data to fill an entire block, resulting in a shorter block
by default. To pad the signal out so that blocks are always
full length, set ``fill_value`` (see below).
frame_length : int > 0
The number of samples per frame.
hop_length : int > 0
The number of samples to advance between frames.
Note that by when ``hop_length < frame_length``, neighboring frames
will overlap. Similarly, the last frame of one *block* will overlap
with the first frame of the next *block*.
mono : bool
Convert the signal to mono during streaming
offset : float
Start reading after this time (in seconds)
duration : float
Only load up to this much audio (in seconds)
fill_value : float [optional]
If padding the signal to produce constant-length blocks,
this value will be used at the end of the signal.
In most cases, ``fill_value=0`` (silence) is expected, but
you may specify any value here.
dtype : numeric type
data type of audio buffers to be produced
Yields
------
y : np.ndarray
An audio buffer of (at most)
``(block_length-1) * hop_length + frame_length`` samples.
See Also
--------
load
get_samplerate
soundfile.blocks
Examples
--------
Apply a short-term Fourier transform to blocks of 256 frames
at a time. Note that streaming operation requires left-aligned
frames, so we must set ``center=False`` to avoid padding artifacts.
>>> filename = librosa.ex('brahms')
>>> sr = librosa.get_samplerate(filename)
>>> stream = librosa.stream(filename,
... block_length=256,
... frame_length=4096,
... hop_length=1024)
>>> for y_block in stream:
... D_block = librosa.stft(y_block, center=False)
Or compute a mel spectrogram over a stream, using a shorter frame
and non-overlapping windows
>>> filename = librosa.ex('brahms')
>>> sr = librosa.get_samplerate(filename)
>>> stream = librosa.stream(filename,
... block_length=256,
... frame_length=2048,
... hop_length=2048)
>>> for y_block in stream:
... m_block = librosa.feature.melspectrogram(y=y_block, sr=sr,
... n_fft=2048,
... hop_length=2048,
... center=False)
"""
if not (np.issubdtype(type(block_length), np.integer) and block_length > 0):
raise ParameterError("block_length={} must be a positive integer")
if not (np.issubdtype(type(frame_length), np.integer) and frame_length > 0):
raise ParameterError("frame_length={} must be a positive integer")
if not (np.issubdtype(type(hop_length), np.integer) and hop_length > 0):
raise ParameterError("hop_length={} must be a positive integer")
if isinstance(path, sf.SoundFile):
sfo = path
else:
sfo = sf.SoundFile(path)
# Get the sample rate from the file info
sr = sfo.samplerate
# Construct the stream
if offset:
start = int(offset * sr)
else:
start = 0
if duration:
frames = int(duration * sr)
else:
frames = -1
# Seek the soundfile object to the starting frame
sfo.seek(start)
blocks = sfo.blocks(
blocksize=frame_length + (block_length - 1) * hop_length,
overlap=frame_length - hop_length,
frames=frames,
dtype=dtype,
always_2d=False,
fill_value=fill_value,
)
for block in blocks:
if mono:
yield to_mono(block.T)
else:
yield block.T
[docs]@cache(level=20)
def to_mono(y):
"""Convert an audio signal to mono by averaging samples across channels.
Parameters
----------
y : np.ndarray [shape=(..., n)]
audio time series. Multi-channel is supported.
Returns
-------
y_mono : np.ndarray [shape=(n,)]
``y`` as a monophonic time-series
Notes
-----
This function caches at level 20.
Examples
--------
>>> y, sr = librosa.load(librosa.ex('trumpet', hq=True), mono=False)
>>> y.shape
(2, 117601)
>>> y_mono = librosa.to_mono(y)
>>> y_mono.shape
(117601,)
"""
# Validate the buffer. Stereo is ok here.
util.valid_audio(y, mono=False)
if y.ndim > 1:
y = np.mean(y, axis=tuple(range(y.ndim - 1)))
return y
[docs]@deprecate_positional_args
@cache(level=20)
def resample(
y, *, orig_sr, target_sr, res_type="kaiser_best", fix=True, scale=False, **kwargs
):
"""Resample a time series from orig_sr to target_sr
By default, this uses a high-quality (but relatively slow) method ('kaiser_best')
for band-limited sinc interpolation. The alternate ``res_type`` values listed below
offer different trade-offs of speed and quality.
Parameters
----------
y : np.ndarray [shape=(..., n)]
audio time series. Multi-channel is supported.
orig_sr : number > 0 [scalar]
original sampling rate of ``y``
target_sr : number > 0 [scalar]
target sampling rate
res_type : str
resample type
'kaiser_best' (default)
`resampy` high-quality mode
'kaiser_fast'
`resampy` faster method
'fft' or 'scipy'
`scipy.signal.resample` Fourier method.
'polyphase'
`scipy.signal.resample_poly` polyphase filtering. (fast)
'linear'
`samplerate` linear interpolation. (very fast)
'zero_order_hold'
`samplerate` repeat the last value between samples. (very fast)
'sinc_best', 'sinc_medium' or 'sinc_fastest'
`samplerate` high-, medium-, and low-quality sinc interpolation.
'soxr_vhq', 'soxr_hq', 'soxr_mq' or 'soxr_lq'
`soxr` Very high-, High-, Medium-, Low-quality FFT-based bandlimited interpolation.
``'soxr_hq'`` is the default setting of `soxr` (fast)
'soxr_qq'
`soxr` Quick cubic interpolation (very fast)
.. note::
`samplerate` and `soxr` are not installed with `librosa`.
To use `samplerate` or `soxr`, they should be installed manually::
$ pip install samplerate
$ pip install soxr
.. note::
When using ``res_type='polyphase'``, only integer sampling rates are
supported.
fix : bool
adjust the length of the resampled signal to be of size exactly
``ceil(target_sr * len(y) / orig_sr)``
scale : bool
Scale the resampled signal so that ``y`` and ``y_hat`` have approximately
equal total energy.
**kwargs : additional keyword arguments
If ``fix==True``, additional keyword arguments to pass to
`librosa.util.fix_length`.
Returns
-------
y_hat : np.ndarray [shape=(..., n * target_sr / orig_sr)]
``y`` resampled from ``orig_sr`` to ``target_sr``
Raises
------
ParameterError
If ``res_type='polyphase'`` and ``orig_sr`` or ``target_sr`` are not both
integer-valued.
See Also
--------
librosa.util.fix_length
scipy.signal.resample
resampy
samplerate.converters.resample
soxr.resample
Notes
-----
This function caches at level 20.
Examples
--------
Downsample from 22 KHz to 8 KHz
>>> y, sr = librosa.load(librosa.ex('trumpet'), sr=22050)
>>> y_8k = librosa.resample(y, orig_sr=sr, target_sr=8000)
>>> y.shape, y_8k.shape
((117601,), (42668,))
"""
# First, validate the audio buffer
util.valid_audio(y, mono=False)
if orig_sr == target_sr:
return y
ratio = float(target_sr) / orig_sr
n_samples = int(np.ceil(y.shape[-1] * ratio))
if res_type in ("scipy", "fft"):
y_hat = scipy.signal.resample(y, n_samples, axis=-1)
elif res_type == "polyphase":
if int(orig_sr) != orig_sr or int(target_sr) != target_sr:
raise ParameterError(
"polyphase resampling is only supported for integer-valued sampling rates."
)
# For polyphase resampling, we need up- and down-sampling ratios
# We can get those from the greatest common divisor of the rates
# as long as the rates are integrable
orig_sr = int(orig_sr)
target_sr = int(target_sr)
gcd = np.gcd(orig_sr, target_sr)
y_hat = scipy.signal.resample_poly(y, target_sr // gcd, orig_sr // gcd, axis=-1)
elif res_type in (
"linear",
"zero_order_hold",
"sinc_best",
"sinc_fastest",
"sinc_medium",
):
import samplerate
# We have to transpose here to match libsamplerate
y_hat = samplerate.resample(y.T, ratio, converter_type=res_type).T
elif res_type.startswith("soxr"):
import soxr
# We have to transpose here to match soxr
y_hat = soxr.resample(y.T, orig_sr, target_sr, quality=res_type).T
else:
y_hat = resampy.resample(y, orig_sr, target_sr, filter=res_type, axis=-1)
if fix:
y_hat = util.fix_length(y_hat, size=n_samples, **kwargs)
if scale:
y_hat /= np.sqrt(ratio)
return y_hat.astype(y.dtype)
[docs]@deprecate_positional_args
def get_duration(
*, y=None, sr=22050, S=None, n_fft=2048, hop_length=512, center=True, filename=None
):
"""Compute the duration (in seconds) of an audio time series,
feature matrix, or filename.
Examples
--------
>>> # Load an example audio file
>>> y, sr = librosa.load(librosa.ex('trumpet'))
>>> librosa.get_duration(y=y, sr=sr)
5.333378684807256
>>> # Or directly from an audio file
>>> librosa.get_duration(filename=librosa.ex('trumpet'))
5.333378684807256
>>> # Or compute duration from an STFT matrix
>>> y, sr = librosa.load(librosa.ex('trumpet'))
>>> S = librosa.stft(y)
>>> librosa.get_duration(S=S, sr=sr)
5.317369614512471
>>> # Or a non-centered STFT matrix
>>> S_left = librosa.stft(y, center=False)
>>> librosa.get_duration(S=S_left, sr=sr)
5.224489795918367
Parameters
----------
y : np.ndarray [shape=(..., n)] or None
audio time series. Multi-channel is supported.
sr : number > 0 [scalar]
audio sampling rate of ``y``
S : np.ndarray [shape=(..., d, t)] or None
STFT matrix, or any STFT-derived matrix (e.g., chromagram
or mel spectrogram).
Durations calculated from spectrogram inputs are only accurate
up to the frame resolution. If high precision is required,
it is better to use the audio time series directly.
n_fft : int > 0 [scalar]
FFT window size for ``S``
hop_length : int > 0 [ scalar]
number of audio samples between columns of ``S``
center : boolean
- If ``True``, ``S[:, t]`` is centered at ``y[t * hop_length]``
- If ``False``, then ``S[:, t]`` begins at ``y[t * hop_length]``
filename : str
If provided, all other parameters are ignored, and the
duration is calculated directly from the audio file.
Note that this avoids loading the contents into memory,
and is therefore useful for querying the duration of
long files.
As in ``load``, this can also be an integer or open file-handle
that can be processed by ``soundfile``.
Returns
-------
d : float >= 0
Duration (in seconds) of the input time series or spectrogram.
Raises
------
ParameterError
if none of ``y``, ``S``, or ``filename`` are provided.
Notes
-----
`get_duration` can be applied to a file (``filename``), a spectrogram (``S``),
or audio buffer (``y, sr``). Only one of these three options should be
provided. If you do provide multiple options (e.g., ``filename`` and ``S``),
then ``filename`` takes precedence over ``S``, and ``S`` takes precedence over
``(y, sr)``.
"""
if filename is not None:
try:
return sf.info(filename).duration
except RuntimeError:
with audioread.audio_open(filename) as fdesc:
return fdesc.duration
if y is None:
if S is None:
raise ParameterError(
"At least one of (y, sr), S, or filename must be provided"
)
n_frames = S.shape[-1]
n_samples = n_fft + hop_length * (n_frames - 1)
# If centered, we lose half a window from each end of S
if center:
n_samples = n_samples - 2 * int(n_fft // 2)
else:
n_samples = y.shape[-1]
return float(n_samples) / sr
[docs]def get_samplerate(path):
"""Get the sampling rate for a given file.
Parameters
----------
path : string, int, soundfile.SoundFile, or file-like
The path to the file to be loaded
As in ``load``, this can also be an integer or open file-handle
that can be processed by `soundfile`.
An existing `soundfile.SoundFile` object can also be supplied.
Returns
-------
sr : number > 0
The sampling rate of the given audio file
Examples
--------
Get the sampling rate for the included audio file
>>> path = librosa.ex('trumpet')
>>> librosa.get_samplerate(path)
22050
"""
try:
if isinstance(path, sf.SoundFile):
return path.samplerate
return sf.info(path).samplerate
except RuntimeError:
with audioread.audio_open(path) as fdesc:
return fdesc.samplerate
[docs]@deprecate_positional_args
@cache(level=20)
def autocorrelate(y, *, max_size=None, axis=-1):
"""Bounded-lag auto-correlation
Parameters
----------
y : np.ndarray
array to autocorrelate
max_size : int > 0 or None
maximum correlation lag.
If unspecified, defaults to ``y.shape[axis]`` (unbounded)
axis : int
The axis along which to autocorrelate.
By default, the last axis (-1) is taken.
Returns
-------
z : np.ndarray
truncated autocorrelation ``y*y`` along the specified axis.
If ``max_size`` is specified, then ``z.shape[axis]`` is bounded
to ``max_size``.
Notes
-----
This function caches at level 20.
Examples
--------
Compute full autocorrelation of ``y``
>>> y, sr = librosa.load(librosa.ex('trumpet'))
>>> librosa.autocorrelate(y)
array([ 6.899e+02, 6.236e+02, ..., 3.710e-08, -1.796e-08])
Compute onset strength auto-correlation up to 4 seconds
>>> import matplotlib.pyplot as plt
>>> odf = librosa.onset.onset_strength(y=y, sr=sr, hop_length=512)
>>> ac = librosa.autocorrelate(odf, max_size=4 * sr // 512)
>>> fig, ax = plt.subplots()
>>> ax.plot(ac)
>>> ax.set(title='Auto-correlation', xlabel='Lag (frames)')
"""
if max_size is None:
max_size = y.shape[axis]
max_size = int(min(max_size, y.shape[axis]))
# Compute the power spectrum along the chosen axis
# Pad out the signal to support full-length auto-correlation.
fft = get_fftlib()
powspec = np.abs(fft.fft(y, n=2 * y.shape[axis] + 1, axis=axis)) ** 2
# Convert back to time domain
autocorr = fft.ifft(powspec, axis=axis)
# Slice down to max_size
subslice = [slice(None)] * autocorr.ndim
subslice[axis] = slice(max_size)
autocorr = autocorr[tuple(subslice)]
if not np.iscomplexobj(y):
autocorr = autocorr.real
return autocorr
[docs]@deprecate_positional_args
def lpc(y, *, order, axis=-1):
"""Linear Prediction Coefficients via Burg's method
This function applies Burg's method to estimate coefficients of a linear
filter on ``y`` of order ``order``. Burg's method is an extension to the
Yule-Walker approach, which are both sometimes referred to as LPC parameter
estimation by autocorrelation.
It follows the description and implementation approach described in the
introduction by Marple. [#]_ N.B. This paper describes a different method, which
is not implemented here, but has been chosen for its clear explanation of
Burg's technique in its introduction.
.. [#] Larry Marple.
A New Autoregressive Spectrum Analysis Algorithm.
IEEE Transactions on Accoustics, Speech, and Signal Processing
vol 28, no. 4, 1980.
Parameters
----------
y : np.ndarray [shape=(..., n)]
Time series to fit. Multi-channel is supported..
order : int > 0
Order of the linear filter
axis : int
Axis along which to compute the coefficients
Returns
-------
a : np.ndarray [shape=(..., order + 1)]
LP prediction error coefficients, i.e. filter denominator polynomial.
Note that the length along the specified ``axis`` will be ``order+1``.
Raises
------
ParameterError
- If ``y`` is not valid audio as per `librosa.util.valid_audio`
- If ``order < 1`` or not integer
FloatingPointError
- If ``y`` is ill-conditioned
See Also
--------
scipy.signal.lfilter
Examples
--------
Compute LP coefficients of y at order 16 on entire series
>>> y, sr = librosa.load(librosa.ex('libri1'))
>>> librosa.lpc(y, order=16)
Compute LP coefficients, and plot LP estimate of original series
>>> import matplotlib.pyplot as plt
>>> import scipy
>>> y, sr = librosa.load(librosa.ex('libri1'), duration=0.020)
>>> a = librosa.lpc(y, order=2)
>>> b = np.hstack([[0], -1 * a[1:]])
>>> y_hat = scipy.signal.lfilter(b, [1], y)
>>> fig, ax = plt.subplots()
>>> ax.plot(y)
>>> ax.plot(y_hat, linestyle='--')
>>> ax.legend(['y', 'y_hat'])
>>> ax.set_title('LP Model Forward Prediction')
"""
if not isinstance(order, (int, np.integer)) or order < 1:
raise ParameterError("order must be an integer > 0")
util.valid_audio(y, mono=False)
# Move the lpc axis around front, because numba is silly
y = y.swapaxes(axis, 0)
dtype = y.dtype
shape = list(y.shape)
shape[0] = order + 1
ar_coeffs = np.zeros(tuple(shape), dtype=dtype)
ar_coeffs[0] = 1
ar_coeffs_prev = ar_coeffs.copy()
shape[0] = 1
reflect_coeff = np.zeros(shape, dtype=dtype)
den = reflect_coeff.copy()
epsilon = util.tiny(den)
# Call the helper, and swap the results back to the target axis position
return __lpc(
y, order, ar_coeffs, ar_coeffs_prev, reflect_coeff, den, epsilon
).swapaxes(0, axis)
@jit(nopython=True, cache=True)
def __lpc(y, order, ar_coeffs, ar_coeffs_prev, reflect_coeff, den, epsilon):
# This implementation follows the description of Burg's algorithm given in
# section III of Marple's paper referenced in the docstring.
#
# We use the Levinson-Durbin recursion to compute AR coefficients for each
# increasing model order by using those from the last. We maintain two
# arrays and then flip them each time we increase the model order so that
# we may use all the coefficients from the previous order while we compute
# those for the new one. These two arrays hold ar_coeffs for order M and
# order M-1. (Corresponding to a_{M,k} and a_{M-1,k} in eqn 5)
# These two arrays hold the forward and backward prediction error. They
# correspond to f_{M-1,k} and b_{M-1,k} in eqns 10, 11, 13 and 14 of
# Marple. First they are used to compute the reflection coefficient at
# order M from M-1 then are re-used as f_{M,k} and b_{M,k} for each
# iteration of the below loop
fwd_pred_error = y[1:]
bwd_pred_error = y[:-1]
# DEN_{M} from eqn 16 of Marple.
den[0] = np.sum(fwd_pred_error ** 2 + bwd_pred_error ** 2, axis=0)
for i in range(order):
# can be removed if we keep the epsilon bias
# if np.any(den <= 0):
# raise FloatingPointError("numerical error, input ill-conditioned?")
# Eqn 15 of Marple, with fwd_pred_error and bwd_pred_error
# corresponding to f_{M-1,k+1} and b{M-1,k} and the result as a_{M,M}
reflect_coeff[0] = np.sum(bwd_pred_error * fwd_pred_error, axis=0)
reflect_coeff[0] *= -2
reflect_coeff[0] /= den[0] + epsilon
# Now we use the reflection coefficient and the AR coefficients from
# the last model order to compute all of the AR coefficients for the
# current one. This is the Levinson-Durbin recursion described in
# eqn 5.
# Note 1: We don't have to care about complex conjugates as our signals
# are all real-valued
# Note 2: j counts 1..order+1, i-j+1 counts order..0
# Note 3: The first element of ar_coeffs* is always 1, which copies in
# the reflection coefficient at the end of the new AR coefficient array
# after the preceding coefficients
ar_coeffs_prev, ar_coeffs = ar_coeffs, ar_coeffs_prev
for j in range(1, i + 2):
# reflection multiply should be broadcast
ar_coeffs[j] = (
ar_coeffs_prev[j] + reflect_coeff[0] * ar_coeffs_prev[i - j + 1]
)
# Update the forward and backward prediction errors corresponding to
# eqns 13 and 14. We start with f_{M-1,k+1} and b_{M-1,k} and use them
# to compute f_{M,k} and b_{M,k}
fwd_pred_error_tmp = fwd_pred_error
fwd_pred_error = fwd_pred_error + reflect_coeff * bwd_pred_error
bwd_pred_error = bwd_pred_error + reflect_coeff * fwd_pred_error_tmp
# SNIP - we are now done with order M and advance. M-1 <- M
# Compute DEN_{M} using the recursion from eqn 17.
#
# reflect_coeff = a_{M-1,M-1} (we have advanced M)
# den = DEN_{M-1} (rhs)
# bwd_pred_error = b_{M-1,N-M+1} (we have advanced M)
# fwd_pred_error = f_{M-1,k} (we have advanced M)
# den <- DEN_{M} (lhs)
#
q = 1.0 - reflect_coeff[0] ** 2
den[0] = q * den[0] - bwd_pred_error[-1] ** 2 - fwd_pred_error[0] ** 2
# Shift up forward error.
#
# fwd_pred_error <- f_{M-1,k+1}
# bwd_pred_error <- b_{M-1,k}
#
# N.B. We do this after computing the denominator using eqn 17 but
# before using it in the numerator in eqn 15.
fwd_pred_error = fwd_pred_error[1:]
bwd_pred_error = bwd_pred_error[:-1]
return ar_coeffs
[docs]@deprecate_positional_args
@cache(level=20)
def zero_crossings(
y, *, threshold=1e-10, ref_magnitude=None, pad=True, zero_pos=True, axis=-1
):
"""Find the zero-crossings of a signal ``y``: indices ``i`` such that
``sign(y[i]) != sign(y[j])``.
If ``y`` is multi-dimensional, then zero-crossings are computed along
the specified ``axis``.
Parameters
----------
y : np.ndarray
The input array
threshold : float > 0 or None
If specified, values where ``-threshold <= y <= threshold`` are
clipped to 0.
ref_magnitude : float > 0 or callable
If numeric, the threshold is scaled relative to ``ref_magnitude``.
If callable, the threshold is scaled relative to
``ref_magnitude(np.abs(y))``.
pad : boolean
If ``True``, then ``y[0]`` is considered a valid zero-crossing.
zero_pos : boolean
If ``True`` then the value 0 is interpreted as having positive sign.
If ``False``, then 0, -1, and +1 all have distinct signs.
axis : int
Axis along which to compute zero-crossings.
Returns
-------
zero_crossings : np.ndarray [shape=y.shape, dtype=boolean]
Indicator array of zero-crossings in ``y`` along the selected axis.
Notes
-----
This function caches at level 20.
Examples
--------
>>> # Generate a time-series
>>> y = np.sin(np.linspace(0, 4 * 2 * np.pi, 20))
>>> y
array([ 0.000e+00, 9.694e-01, 4.759e-01, -7.357e-01,
-8.372e-01, 3.247e-01, 9.966e-01, 1.646e-01,
-9.158e-01, -6.142e-01, 6.142e-01, 9.158e-01,
-1.646e-01, -9.966e-01, -3.247e-01, 8.372e-01,
7.357e-01, -4.759e-01, -9.694e-01, -9.797e-16])
>>> # Compute zero-crossings
>>> z = librosa.zero_crossings(y)
>>> z
array([ True, False, False, True, False, True, False, False,
True, False, True, False, True, False, False, True,
False, True, False, True], dtype=bool)
>>> # Stack y against the zero-crossing indicator
>>> librosa.util.stack([y, z], axis=-1)
array([[ 0.000e+00, 1.000e+00],
[ 9.694e-01, 0.000e+00],
[ 4.759e-01, 0.000e+00],
[ -7.357e-01, 1.000e+00],
[ -8.372e-01, 0.000e+00],
[ 3.247e-01, 1.000e+00],
[ 9.966e-01, 0.000e+00],
[ 1.646e-01, 0.000e+00],
[ -9.158e-01, 1.000e+00],
[ -6.142e-01, 0.000e+00],
[ 6.142e-01, 1.000e+00],
[ 9.158e-01, 0.000e+00],
[ -1.646e-01, 1.000e+00],
[ -9.966e-01, 0.000e+00],
[ -3.247e-01, 0.000e+00],
[ 8.372e-01, 1.000e+00],
[ 7.357e-01, 0.000e+00],
[ -4.759e-01, 1.000e+00],
[ -9.694e-01, 0.000e+00],
[ -9.797e-16, 1.000e+00]])
>>> # Find the indices of zero-crossings
>>> np.nonzero(z)
(array([ 0, 3, 5, 8, 10, 12, 15, 17, 19]),)
"""
# Clip within the threshold
if threshold is None:
threshold = 0.0
if callable(ref_magnitude):
threshold = threshold * ref_magnitude(np.abs(y))
elif ref_magnitude is not None:
threshold = threshold * ref_magnitude
if threshold > 0:
y = y.copy()
y[np.abs(y) <= threshold] = 0
# Extract the sign bit
if zero_pos:
y_sign = np.signbit(y)
else:
y_sign = np.sign(y)
# Find the change-points by slicing
slice_pre = [slice(None)] * y.ndim
slice_pre[axis] = slice(1, None)
slice_post = [slice(None)] * y.ndim
slice_post[axis] = slice(-1)
# Since we've offset the input by one, pad back onto the front
padding = [(0, 0)] * y.ndim
padding[axis] = (1, 0)
return np.pad(
(y_sign[tuple(slice_post)] != y_sign[tuple(slice_pre)]),
padding,
mode="constant",
constant_values=pad,
)
[docs]@deprecate_positional_args
def clicks(
*,
times=None,
frames=None,
sr=22050,
hop_length=512,
click_freq=1000.0,
click_duration=0.1,
click=None,
length=None,
):
"""Construct a "click track".
This returns a signal with the signal ``click`` sound placed at
each specified time.
Parameters
----------
times : np.ndarray or None
times to place clicks, in seconds
frames : np.ndarray or None
frame indices to place clicks
sr : number > 0
desired sampling rate of the output signal
hop_length : int > 0
if positions are specified by ``frames``, the number of samples between frames.
click_freq : float > 0
frequency (in Hz) of the default click signal. Default is 1KHz.
click_duration : float > 0
duration (in seconds) of the default click signal. Default is 100ms.
click : np.ndarray or None
(optional) click signal sample to use instead of the default click.
Multi-channel is supported.
length : int > 0
desired number of samples in the output signal
Returns
-------
click_signal : np.ndarray
Synthesized click signal.
This will be monophonic by default, or match the number of channels to a provided ``click`` signal.
Raises
------
ParameterError
- If neither ``times`` nor ``frames`` are provided.
- If any of ``click_freq``, ``click_duration``, or ``length`` are out of range.
Examples
--------
>>> # Sonify detected beat events
>>> y, sr = librosa.load(librosa.ex('choice'), duration=10)
>>> tempo, beats = librosa.beat.beat_track(y=y, sr=sr)
>>> y_beats = librosa.clicks(frames=beats, sr=sr)
>>> # Or generate a signal of the same length as y
>>> y_beats = librosa.clicks(frames=beats, sr=sr, length=len(y))
>>> # Or use timing instead of frame indices
>>> times = librosa.frames_to_time(beats, sr=sr)
>>> y_beat_times = librosa.clicks(times=times, sr=sr)
>>> # Or with a click frequency of 880Hz and a 500ms sample
>>> y_beat_times880 = librosa.clicks(times=times, sr=sr,
... click_freq=880, click_duration=0.5)
Display click waveform next to the spectrogram
>>> import matplotlib.pyplot as plt
>>> fig, ax = plt.subplots(nrows=2, sharex=True)
>>> S = librosa.feature.melspectrogram(y=y, sr=sr)
>>> librosa.display.specshow(librosa.power_to_db(S, ref=np.max),
... x_axis='time', y_axis='mel', ax=ax[0])
>>> librosa.display.waveshow(y_beat_times, sr=sr, label='Beat clicks',
... ax=ax[1])
>>> ax[1].legend()
>>> ax[0].label_outer()
>>> ax[0].set_title(None)
"""
# Compute sample positions from time or frames
if times is None:
if frames is None:
raise ParameterError('either "times" or "frames" must be provided')
positions = frames_to_samples(frames, hop_length=hop_length)
else:
# Convert times to positions
positions = time_to_samples(times, sr=sr)
if click is not None:
# Check that we have a well-formed audio buffer
util.valid_audio(click, mono=False)
else:
# Create default click signal
if click_duration <= 0:
raise ParameterError("click_duration must be strictly positive")
if click_freq <= 0:
raise ParameterError("click_freq must be strictly positive")
angular_freq = 2 * np.pi * click_freq / float(sr)
click = np.logspace(0, -10, num=int(np.round(sr * click_duration)), base=2.0)
click *= np.sin(angular_freq * np.arange(len(click)))
# Set default length
if length is None:
length = positions.max() + click.shape[-1]
else:
if length < 1:
raise ParameterError("length must be a positive integer")
# Filter out any positions past the length boundary
positions = positions[positions < length]
# Pre-allocate click signal
shape = list(click.shape)
shape[-1] = length
click_signal = np.zeros(shape, dtype=np.float32)
# Place clicks
for start in positions:
# Compute the end-point of this click
end = start + click.shape[-1]
if end >= length:
click_signal[..., start:] += click[..., : length - start]
else:
# Normally, just add a click here
click_signal[..., start:end] += click
return click_signal
[docs]@deprecate_positional_args
def tone(frequency, *, sr=22050, length=None, duration=None, phi=None):
"""Construct a pure tone (cosine) signal at a given frequency.
Parameters
----------
frequency : float > 0
frequency
sr : number > 0
desired sampling rate of the output signal
length : int > 0
desired number of samples in the output signal.
When both ``duration`` and ``length`` are defined,
``length`` takes priority.
duration : float > 0
desired duration in seconds.
When both ``duration`` and ``length`` are defined,
``length`` takes priority.
phi : float or None
phase offset, in radians. If unspecified, defaults to ``-np.pi * 0.5``.
Returns
-------
tone_signal : np.ndarray [shape=(length,), dtype=float64]
Synthesized pure sine tone signal
Raises
------
ParameterError
- If ``frequency`` is not provided.
- If neither ``length`` nor ``duration`` are provided.
Examples
--------
Generate a pure sine tone A4
>>> tone = librosa.tone(440, duration=1)
Or generate the same signal using `length`
>>> tone = librosa.tone(440, sr=22050, length=22050)
Display spectrogram
>>> import matplotlib.pyplot as plt
>>> fig, ax = plt.subplots()
>>> S = librosa.feature.melspectrogram(y=tone)
>>> librosa.display.specshow(librosa.power_to_db(S, ref=np.max),
... x_axis='time', y_axis='mel', ax=ax)
"""
if frequency is None:
raise ParameterError('"frequency" must be provided')
# Compute signal length
if length is None:
if duration is None:
raise ParameterError('either "length" or "duration" must be provided')
length = duration * sr
if phi is None:
phi = -np.pi * 0.5
return np.cos(2 * np.pi * frequency * np.arange(length) / sr + phi)
[docs]@deprecate_positional_args
def chirp(*, fmin, fmax, sr=22050, length=None, duration=None, linear=False, phi=None):
"""Construct a "chirp" or "sine-sweep" signal.
The chirp sweeps from frequency ``fmin`` to ``fmax`` (in Hz).
Parameters
----------
fmin : float > 0
initial frequency
fmax : float > 0
final frequency
sr : number > 0
desired sampling rate of the output signal
length : int > 0
desired number of samples in the output signal.
When both ``duration`` and ``length`` are defined,
``length`` takes priority.
duration : float > 0
desired duration in seconds.
When both ``duration`` and ``length`` are defined,
``length`` takes priority.
linear : boolean
- If ``True``, use a linear sweep, i.e., frequency changes linearly with time
- If ``False``, use a exponential sweep.
Default is ``False``.
phi : float or None
phase offset, in radians.
If unspecified, defaults to ``-np.pi * 0.5``.
Returns
-------
chirp_signal : np.ndarray [shape=(length,), dtype=float64]
Synthesized chirp signal
Raises
------
ParameterError
- If either ``fmin`` or ``fmax`` are not provided.
- If neither ``length`` nor ``duration`` are provided.
See Also
--------
scipy.signal.chirp
Examples
--------
Generate a exponential chirp from A2 to A8
>>> exponential_chirp = librosa.chirp(fmin=110, fmax=110*64, duration=1)
Or generate the same signal using ``length``
>>> exponential_chirp = librosa.chirp(fmin=110, fmax=110*64, sr=22050, length=22050)
Or generate a linear chirp instead
>>> linear_chirp = librosa.chirp(fmin=110, fmax=110*64, duration=1, linear=True)
Display spectrogram for both exponential and linear chirps.
>>> import matplotlib.pyplot as plt
>>> fig, ax = plt.subplots(nrows=2, sharex=True, sharey=True)
>>> S_exponential = np.abs(librosa.stft(y=exponential_chirp))
>>> librosa.display.specshow(librosa.amplitude_to_db(S_exponential, ref=np.max),
... x_axis='time', y_axis='linear', ax=ax[0])
>>> ax[0].set(title='Exponential chirp', xlabel=None)
>>> ax[0].label_outer()
>>> S_linear = np.abs(librosa.stft(y=linear_chirp))
>>> librosa.display.specshow(librosa.amplitude_to_db(S_linear, ref=np.max),
... x_axis='time', y_axis='linear', ax=ax[1])
>>> ax[1].set(title='Linear chirp')
"""
if fmin is None or fmax is None:
raise ParameterError('both "fmin" and "fmax" must be provided')
# Compute signal duration
period = 1.0 / sr
if length is None:
if duration is None:
raise ParameterError('either "length" or "duration" must be provided')
else:
duration = period * length
if phi is None:
phi = -np.pi * 0.5
method = "linear" if linear else "logarithmic"
return scipy.signal.chirp(
np.arange(duration, step=period),
fmin,
duration,
fmax,
method=method,
phi=phi / np.pi * 180, # scipy.signal.chirp uses degrees for phase offset
)
[docs]@deprecate_positional_args
def mu_compress(x, *, mu=255, quantize=True):
"""mu-law compression
Given an input signal ``-1 <= x <= 1``, the mu-law compression
is calculated by::
sign(x) * ln(1 + mu * abs(x)) / ln(1 + mu)
Parameters
----------
x : np.ndarray with values in [-1, +1]
The input signal to compress
mu : positive number
The compression parameter. Values of the form ``2**n - 1``
(e.g., 15, 31, 63, etc.) are most common.
quantize : bool
If ``True``, quantize the compressed values into ``1 + mu``
distinct integer values.
If ``False``, mu-law compression is applied without quantization.
Returns
-------
x_compressed : np.ndarray
The compressed signal.
Raises
------
ParameterError
If ``x`` has values outside the range [-1, +1]
If ``mu <= 0``
See Also
--------
mu_expand
Examples
--------
Compression without quantization
>>> x = np.linspace(-1, 1, num=16)
>>> x
array([-1. , -0.86666667, -0.73333333, -0.6 , -0.46666667,
-0.33333333, -0.2 , -0.06666667, 0.06666667, 0.2 ,
0.33333333, 0.46666667, 0.6 , 0.73333333, 0.86666667,
1. ])
>>> y = librosa.mu_compress(x, quantize=False)
>>> y
array([-1. , -0.97430198, -0.94432361, -0.90834832, -0.86336132,
-0.80328309, -0.71255496, -0.52124063, 0.52124063, 0.71255496,
0.80328309, 0.86336132, 0.90834832, 0.94432361, 0.97430198,
1. ])
Compression with quantization
>>> y = librosa.mu_compress(x, quantize=True)
>>> y
array([-128, -124, -120, -116, -110, -102, -91, -66, 66, 91, 102,
110, 116, 120, 124, 127])
Compression with quantization and a smaller range
>>> y = librosa.mu_compress(x, mu=15, quantize=True)
>>> y
array([-8, -7, -7, -6, -6, -5, -4, -2, 2, 4, 5, 6, 6, 7, 7, 7])
"""
if mu <= 0:
raise ParameterError(
"mu-law compression parameter mu={} "
"must be strictly positive.".format(mu)
)
if np.any(x < -1) or np.any(x > 1):
raise ParameterError(
"mu-law input x={} must be in the " "range [-1, +1].".format(x)
)
x_comp = np.sign(x) * np.log1p(mu * np.abs(x)) / np.log1p(mu)
if quantize:
return (
np.digitize(
x_comp, np.linspace(-1, 1, num=int(1 + mu), endpoint=True), right=True
)
- int(mu + 1) // 2
)
return x_comp
[docs]@deprecate_positional_args
def mu_expand(x, *, mu=255.0, quantize=True):
"""mu-law expansion
This function is the inverse of ``mu_compress``. Given a mu-law compressed
signal ``-1 <= x <= 1``, the mu-law expansion is calculated by::
sign(x) * (1 / mu) * ((1 + mu)**abs(x) - 1)
Parameters
----------
x : np.ndarray
The compressed signal.
If ``quantize=True``, values must be in the range [-1, +1].
mu : positive number
The compression parameter. Values of the form ``2**n - 1``
(e.g., 15, 31, 63, etc.) are most common.
quantize : boolean
If ``True``, the input is assumed to be quantized to
``1 + mu`` distinct integer values.
Returns
-------
x_expanded : np.ndarray with values in the range [-1, +1]
The mu-law expanded signal.
Raises
------
ParameterError
If ``x`` has values outside the range [-1, +1] and ``quantize=False``
If ``mu <= 0``
See Also
--------
mu_compress
Examples
--------
Compress and expand without quantization
>>> x = np.linspace(-1, 1, num=16)
>>> x
array([-1. , -0.86666667, -0.73333333, -0.6 , -0.46666667,
-0.33333333, -0.2 , -0.06666667, 0.06666667, 0.2 ,
0.33333333, 0.46666667, 0.6 , 0.73333333, 0.86666667,
1. ])
>>> y = librosa.mu_compress(x, quantize=False)
>>> y
array([-1. , -0.97430198, -0.94432361, -0.90834832, -0.86336132,
-0.80328309, -0.71255496, -0.52124063, 0.52124063, 0.71255496,
0.80328309, 0.86336132, 0.90834832, 0.94432361, 0.97430198,
1. ])
>>> z = librosa.mu_expand(y, quantize=False)
>>> z
array([-1. , -0.86666667, -0.73333333, -0.6 , -0.46666667,
-0.33333333, -0.2 , -0.06666667, 0.06666667, 0.2 ,
0.33333333, 0.46666667, 0.6 , 0.73333333, 0.86666667,
1. ])
Compress and expand with quantization. Note that this necessarily
incurs quantization error, particularly for values near +-1.
>>> y = librosa.mu_compress(x, quantize=True)
>>> y
array([-128, -124, -120, -116, -110, -102, -91, -66, 66, 91, 102,
110, 116, 120, 124, 127])
>>> z = librosa.mu_expand(y, quantize=True)
array([-1. , -0.84027248, -0.70595818, -0.59301377, -0.4563785 ,
-0.32155973, -0.19817918, -0.06450245, 0.06450245, 0.19817918,
0.32155973, 0.4563785 , 0.59301377, 0.70595818, 0.84027248,
0.95743702])
"""
if mu <= 0:
raise ParameterError(
"Inverse mu-law compression parameter "
"mu={} must be strictly positive.".format(mu)
)
if quantize:
x = x * 2.0 / (1 + mu)
if np.any(x < -1) or np.any(x > 1):
raise ParameterError(
"Inverse mu-law input x={} must be " "in the range [-1, +1].".format(x)
)
return np.sign(x) / mu * (np.power(1 + mu, np.abs(x)) - 1)