Caution
You're reading the documentation for a development version. For the latest released version, please have a look at 0.11.0.
Source code for librosa.core.audio
#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""Core IO, DSP and utility functions."""
from __future__ import annotations
from typing import TYPE_CHECKING, overload
import lazy_loader as lazy
import numpy as np
import scipy
import soundfile as sf
import soxr
from numba import guvectorize, jit, stencil
from .. import util
from .._cache import cache
from ..util.decorators import future_default
from ..util.exceptions import ParameterError
from ..util.files import example
from .convert import frames_to_samples, time_to_samples
if TYPE_CHECKING:
import os
from typing import Any, BinaryIO, Callable, Generator
from numpy.typing import DTypeLike, NDArray
from .._typing import _Array1D, _FloatLike_co, _IntLike_co, _SequenceLike
# Lazy-load optional dependencies
samplerate = lazy.load("samplerate")
resampy = lazy.load("resampy")
__all__ = [
"load",
"loadx",
"stream",
"to_mono",
"to_stereo",
"to_multi",
"resample",
"get_duration",
"get_samplerate",
"autocorrelate",
"lpc",
"zero_crossings",
"clicks",
"tone",
"chirp",
"mu_compress",
"mu_expand",
]
# -- CORE ROUTINES --#
# Load should never be cached, since we cannot verify that the contents of
# 'path' are unchanged across calls.
[docs]
def load(
path: str | int | os.PathLike[Any] | sf.SoundFile | BinaryIO,
*,
sr: float | None = 22050,
mono: bool = True,
offset: float = 0.0,
duration: float | None = None,
dtype: DTypeLike = np.float32,
res_type: str = "soxr_hq",
) -> tuple[np.ndarray, int | float]:
"""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 : str, int, pathlib.Path, soundfile.SoundFile, or file-like object
path to the input file.
Any codec supported by `soundfile` 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.
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).
If negative, it will be interpreted relative to the end of the file.
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 `soxr`'s high-quality mode ('HQ').
For alternative resampling modes, see `resample`
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)
"""
y, sr_native = __soundfile_load(path, offset, duration, dtype)
# 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 != 0:
if offset > 0:
# Seek to the start of the target read
sf_desc.seek(int(offset * sr_native))
else:
sf_desc.seek(-int(abs(offset) * sr_native), whence=sf.SEEK_END)
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 _align_step_size(target_step, target_sr, orig_sr):
"""
When resampling a stream, we need to ensure that the target step size
remains integer-valued in the original sampling rate
"""
if target_sr is None or target_sr == orig_sr:
return target_step
exact_native_step = (target_step * orig_sr) / target_sr
aligned_native_step = np.round(exact_native_step)
# Tolerances (rtol=1e-7, atol=1e-5) safely absorb precision errors from float arithmetic
if not np.isclose(exact_native_step, aligned_native_step, rtol=1e-7, atol=1e-5):
raise ParameterError(
f"Target block step of {target_step} samples at {target_sr} results in a "
f"fractional sample advance at original sampling rate {orig_sr}."
)
return int(aligned_native_step)
[docs]
@future_default(param_name="sr", old_default=None, new_default=22050, version="1.1.0")
def stream(
path: str | int | sf.SoundFile | BinaryIO,
*,
block_length: int,
frame_length: int,
hop_length: int,
sr: float | None = None,
mono: bool = True,
offset: float = 0.0,
duration: float | None = None,
fill_value: float | None = None,
res_type: str = "soxr_hq",
dtype: DTypeLike = np.float32,
) -> Generator[np.ndarray, None, None]:
"""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 supported,
but not all combinations of ``block_length``, ``frame_length``, and
``hop_length`` will be compatible with all sampling rates.
Specifically, ``(block_length * hop_length * native_sr) / sr``
must evaluate to an exact integer, where ``native_sr`` is the original
sampling rate of the stream prior to resampling.
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.
6. If you break out of the generator loop early, the underlying audio
file handle and resampling buffers will remain open until the
generator object is garbage-collected. To explicitly release resources,
call the generator's ``.close()`` method.
See the examples below for proper usage of this function.
Parameters
----------
path : str, 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*.
sr : number > 0 [scalar]
target sampling rate. If not provided, the original sampling rate of the file will be
used.
mono : bool
Convert the signal to mono during streaming
offset : float
Start reading after this time (in seconds)
If negative, it will be interpreted relative to the end of the file.
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.
res_type : str
Resample type, must be one of the following:
'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`.
'soxr_qq'
`soxr` Quick cubic interpolation (very fast, but not bandlimited)
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 util.is_positive_int(block_length):
raise ParameterError(f"block_length={block_length} must be a positive integer")
if not util.is_positive_int(frame_length):
raise ParameterError(f"frame_length={frame_length} must be a positive integer")
if not util.is_positive_int(hop_length):
raise ParameterError(f"hop_length={hop_length} must be a positive integer")
if sr is not None and not (np.isfinite(sr) and sr > 0):
raise ParameterError(f"sr={sr} must be a positive number")
if res_type not in {"soxr_vhq", "soxr_hq", "soxr_mq", "soxr_lq", "soxr_qq"}:
raise ParameterError(f"res_type={res_type} is not a valid soxr resampling mode for streaming")
if isinstance(path, sf.SoundFile):
sfo = path
else:
sfo = sf.SoundFile(path)
# Get the sample rate from the file info
orig_sr = sfo.samplerate
# Construct the stream
# Seek the soundfile object to the starting frame
target_yield_size = (block_length - 1) * hop_length + frame_length
target_advance = block_length * hop_length
# Determine internal channel state for buffer initialization
is_multichannel = sfo.channels > 1
process_channels = 1 if mono else sfo.channels
needs_resampling = (sr is not None) and (orig_sr != sr)
if sr is None:
sr = orig_sr
try:
orig_read_size = _align_step_size(target_advance, sr, orig_sr)
read_frames = int(duration * orig_sr) if duration is not None else -1
if needs_resampling:
resampler = soxr.ResampleStream(
in_rate=orig_sr,
out_rate=sr,
num_channels=process_channels,
dtype=dtype,
quality=res_type,
)
capacity = target_yield_size + (target_advance * 2)
buffer_shape = (capacity,) if process_channels == 1 else (capacity, process_channels)
buffer = np.zeros(buffer_shape, dtype=dtype)
write_idx = 0
read_idx = 0
if offset >= 0:
sfo.seek(int(offset * orig_sr))
else:
sfo.seek(-int(abs(offset) * orig_sr), whence=sf.SEEK_END)
# Reusable buffer for the per-chunk mono downmix; the final short
# chunk writes into a truncated view of it, avoiding a fresh
# allocation on every block
mono_buffer = (
np.empty(orig_read_size, dtype=dtype)
if mono and is_multichannel
else None
)
for orig_chunk in sfo.blocks(blocksize=orig_read_size,
overlap=0,
dtype=dtype,
always_2d=False,
frames=read_frames):
if mono and is_multichannel:
# soundfile returns (samples, channels), to_mono expects (channels, samples)
assert mono_buffer is not None # for mypy; set above under same condition
orig_chunk = to_mono(
orig_chunk.T, out=mono_buffer[: orig_chunk.shape[0]]
)
if needs_resampling:
target_chunk = resampler.resample_chunk(orig_chunk)
else:
target_chunk = orig_chunk
n_incoming = target_chunk.shape[0]
if write_idx + n_incoming > capacity:
available = write_idx - read_idx
buffer[:available] = buffer[read_idx : write_idx]
read_idx = 0
write_idx = available
# The following branch should never happen, but it's here
# just in case something goes wrong with the input stream
if write_idx + n_incoming > capacity: # pragma: no cover
capacity = write_idx + n_incoming + target_yield_size
new_shape = (capacity,) if process_channels == 1 else (capacity, process_channels)
new_buffer = np.zeros(new_shape, dtype=dtype)
new_buffer[:write_idx] = buffer[:write_idx]
buffer = new_buffer
buffer[write_idx : write_idx + n_incoming] = target_chunk
write_idx += n_incoming
while write_idx - read_idx >= target_yield_size:
block = buffer[read_idx : read_idx + target_yield_size]
yield block.T.copy()
read_idx += target_advance
if process_channels == 1:
empty = np.empty((0,), dtype=buffer.dtype)
else:
empty = np.empty((0, process_channels), dtype=buffer.dtype)
if needs_resampling:
tail_chunk = resampler.resample_chunk(empty, last=True)
else:
tail_chunk = None
final_data_list = [buffer[read_idx : write_idx]]
if tail_chunk is not None and tail_chunk.shape[0] > 0:
final_data_list.append(tail_chunk)
remainder = np.concatenate(final_data_list, axis=0) if final_data_list else np.array([])
else:
remainder = buffer[read_idx : write_idx]
rem_idx = 0
while rem_idx < remainder.shape[0]:
current_slice = remainder[rem_idx : rem_idx + target_yield_size]
if current_slice.shape[0] < target_yield_size:
pad_length = target_yield_size - current_slice.shape[0]
pad_width = (0, pad_length) if process_channels == 1 else ((0, pad_length), (0, 0))
if fill_value is not None:
current_slice = np.pad(current_slice,
pad_width,
mode="constant",
constant_values=fill_value)
yield current_slice.T.copy()
rem_idx += target_advance
finally:
# If we instantiated a new SoundFile object, we should close it. If the
# user passed in an existing SoundFile object, we should leave it alone.
if not isinstance(path, sf.SoundFile):
sfo.close()
[docs]
def loadx(key: str, *, hq: bool | None = None, **kwargs: Any) -> tuple[np.ndarray, int | float]:
"""Load an example audio file by key.
This is a wrapper around `librosa.util.example` that provides the same
functionality as `load`, but with the convenience of loading from the
built-in examples.
.. note:: This function is primarily useful for demonstration purposes,
and to make documentation examples more concise.
For general-purpose loading of audio files, use `load` instead.
Parameters
----------
key : str
The identifier for the track to load
hq : bool, optional
If ``True``, return the high-quality version of the recording.
If ``False``, return the 22KHz mono version of the recording.
If not provided, `hq` is inferred based on additional parameters in `kwargs`:
- If `sr` is provided and greater than 22050 (or is `None`), then `hq` is set to `True`.
- If `mono` is provided and set to `False`, then `hq` is set to `True`.
- Otherwise, `hq` is set to `False`.
**kwargs : additional keyword arguments
Additional keyword arguments to pass to `load`
Returns
-------
y : np.ndarray [shape=(n,) or (..., n)]
audio time series. Multi-channel is supported.
sr : number > 0 [scalar]
sampling rate of ``y``
See Also
--------
load
librosa.util.example
librosa.util.list_examples
Examples
--------
Load the default version of the 'trumpet' example
>>> y, sr = librosa.loadx('trumpet')
Load the stereo version of the 'trumpet' example
>>> y, sr = librosa.loadx('trumpet', mono=False)
>>> # This is equivalent to the following more verbose code:
>>> y, sr = librosa.load(librosa.ex('trumpet', hq=True), mono=False)
"""
if hq is None:
if (("sr" in kwargs and (kwargs["sr"] is None or kwargs["sr"] > 22050)) or
("mono" in kwargs and kwargs["mono"] is False)
):
hq = True
else:
hq = False
try:
path = example(key, hq=hq)
except ParameterError as exc:
raise ParameterError(f"Could not load example with key '{key}'. "
"Did you mean to use librosa.load instead of loadx?") from exc
return load(path, **kwargs)
def _init_output(
out: np.ndarray | None, shape: tuple[int, ...], dtype: np.dtype
) -> np.ndarray:
"""Validate a caller-provided output buffer, or allocate a new one.
If ``out`` is ``None``, return a new zero-filled array of the given shape
and dtype. Otherwise ``out`` must match ``shape`` exactly, and its dtype
must be able to hold ``dtype`` without loss of precision (i.e. at least as
precise as ``dtype``); it is zeroed in place and returned. A caller that
explicitly allocates a higher-precision buffer keeps that precision.
"""
if out is None:
return np.zeros(shape, dtype=dtype)
if out.shape != shape:
raise ParameterError(f"out has shape={out.shape}, but expected {shape}")
if np.promote_types(out.dtype, dtype) != out.dtype:
raise ParameterError(
f"out has dtype={out.dtype}, which cannot hold the result dtype {dtype}"
)
out[...] = 0
return out
[docs]
def to_mono(
*signals: np.ndarray,
pad: bool = True,
norm: bool = True,
out: np.ndarray | None = None,
) -> np.ndarray:
"""Convert an audio signal to mono by averaging samples across channels.
Parameters
----------
*signals : np.ndarray [shape=(..., n)]
One or more signals to convert to mono.
Each input signal can be mono or multi-channel, and they need not all have
identical lengths.
pad : bool
If `True`, pad the output to match the length of the longest input:
`n_out = max(y.shape[-1] for y in signals)`.
If `False`, trim the output to match the length of the shortest input:
`n_out = min(y.shape[-1] for y in signals)`.
norm : bool
If `True` (default), signals are combined by averaging.
If `False`, signals are combined by summing.
out : np.ndarray [shape=(n_out,)] or None
A pre-allocated output buffer. If provided, it must match the shape
``(n_out,)``, and its dtype must be at least as precise as the result
dtype; it is zeroed and filled in place and returned, avoiding a new
allocation. If ``None`` (default), a new array is allocated.
Returns
-------
y_mono : np.ndarray [shape=(n_out,)]
All signals combined together into a single mono signal.
If ``out`` was provided, this is the same object as ``out``.
Notes
-----
The dtype of the output signal will be the most general type that can
accommodate all input signals. If the input signals have different
dtypes, the output will be promoted to the most general type.
.. warning:: Any input signal with more than one channel will be
downmixed to mono prior to being combined with other signals.
This means that the following are generally not equivalent
for multi-channel inputs `y1` and `y2` when `norm=True`:
>>> y_mono = librosa.to_mono(y1, y2)
>>> y_mono = librosa.to_mono(np.vstack((y1, y2)))
See Also
--------
to_stereo
to_multi
util.fix_length
Examples
--------
Downmix a stereo input
>>> y, sr = librosa.loadx('trumpet', mono=False)
>>> y.shape
(2, 117601)
>>> y_mono = librosa.to_mono(y)
>>> y_mono.shape
(117601,)
Mix three independent mono inputs with different lengths
>>> y_440 = librosa.tone(440.0, sr=22050, duration=1.0)
>>> y_550 = librosa.tone(550.0, sr=22050, duration=1.5)
>>> y_660 = librosa.tone(660.0, sr=22050, duration=2.0)
>>> y_mono = librosa.to_mono(y_440, y_550, y_660, pad=True)
>>> y_mono.shape
(44100,)
"""
# Validate the buffer(s) and identify lengths, dtypes, etc
if not signals:
raise ParameterError("At least one signal must be provided to `to_mono`.")
n_min, n_max = signals[0].shape[-1], signals[0].shape[-1]
dtype = signals[0].dtype
for y in signals:
util.valid_audio(y)
n_min = min(n_min, y.shape[-1])
n_max = max(n_max, y.shape[-1])
dtype = np.promote_types(dtype, y.dtype)
# Allocate and initialize the output buffer
if pad:
size = n_max
else:
size = n_min
output = _init_output(out, (size,), dtype)
if norm:
combine = np.mean
else:
combine = np.sum
# Now, average the signals together
for y in signals:
output += util.fix_length(
combine(y, axis=tuple(range(y.ndim - 1))), size=output.shape[-1], axis=-1
)
if norm:
# Divide by the number of input signals given
output /= len(signals)
return output
[docs]
def to_stereo(
*,
left: np.ndarray | None,
right: np.ndarray | None,
downmix: bool = True,
pad: bool = True,
norm: bool = True,
out: np.ndarray | None = None,
) -> np.ndarray:
"""Combine two signals into a stereo signal.
Parameters
----------
left : np.ndarray [shape=(..., n)] or None
Left channel signal. Multi-channel is supported.
right : np.ndarray [shape=(..., m)] or None
Right channel signal. Multi-channel is supported.
downmix : bool
If `True`, downmix the left and right channels to mono before combining.
If `False`, the left and right signals are additively combined.
pad : bool
If `True`, pad the shorter channel with zeros to match the length of the longer channel.
If `False`, the longer channel is trimmed to match the length of the shorter channel.
norm : bool
If `True` (default), signals are combined by averaging.
If `False`, signals are combined by summing.
out : np.ndarray [shape=(2, n_out)] or None
A pre-allocated output buffer. If provided, it must match the shape
``(2, n_out)``, and its dtype must be at least as precise as the result
dtype; it is zeroed and filled in place and returned, avoiding a new
allocation. If ``None`` (default), a new array is allocated.
Returns
-------
y_stereo : np.ndarray [shape=(2, n_out)]
Stereo signal with left channel in the first row and right channel in the second row.
If ``out`` was provided, this is the same object as ``out``.
See Also
--------
to_mono
to_multi
util.fix_length
Notes
-----
At least one of `left` or `right` must be provided.
Examples
--------
Combine two mono signals into a hard-panned stereo signal
>>> y1 = librosa.tone(440.0, sr=22050, duration=1.0)
>>> y2 = librosa.tone(550.0, sr=22050, duration=1.0)
>>> y_stereo = librosa.to_stereo(left=y1, right=y2)
Upmix a mono signal into the left or right channel in stereo
>>> y_left = librosa.to_stereo(left=y1, right=None)
>>> y_right = librosa.to_stereo(left=None, right=y2)
Downmix a stereo signal into the left channel, and mix a third
mono signal into the right channel
>>> y3 = librosa.tone(660.0, sr=22050, duration=1.0)
>>> y_mix = librosa.to_stereo(left=y_stereo, right=y3, downmix=True)
Keep the original stereo signal separated, but mix a third signal into the
right channel:
>>> y_mix = librosa.to_stereo(left=y_stereo, right=y3, downmix=False)
"""
# This flag tracks whether we had only one signal to begin with
onesided = True
if left is None and right is None:
raise ParameterError("At least one of 'left' or 'right' must be provided")
elif left is None:
# These are somewhat inefficient ways to allocate a silent channel,
# but it makes the logic simple and clear below
left = np.zeros_like(right)
elif right is None:
right = np.zeros_like(left)
else:
onesided = False
# This assert tells mypy that both arrays now exist
assert left is not None and right is not None
# First, deal with padding
if pad:
size = max(left.shape[-1], right.shape[-1])
else:
size = min(left.shape[-1], right.shape[-1])
# This implements either padding or trimming
left = util.fix_length(left, size=size, axis=-1)
right = util.fix_length(right, size=size, axis=-1)
# Get a compatible dtype for both channels
dtype = np.promote_types(left.dtype, right.dtype)
left = left.astype(dtype, copy=False)
right = right.astype(dtype, copy=False)
# Create an empty stereo output buffer
output = _init_output(out, (2, size), dtype)
if downmix:
# Downmix directly into the output rows to avoid a second allocation
to_mono(left, norm=norm, out=output[0])
to_mono(right, norm=norm, out=output[1])
else:
if left.ndim == 1:
output[0] = left
elif left.ndim == 2 and left.shape[0] == 2:
output[:] = left
else:
raise ParameterError(
f"left input has unsupported shape {left.shape} for downmix=False"
)
if right.ndim == 1:
output[1] += right
elif right.ndim == 2 and right.shape[0] == 2:
output[:] += right
else:
raise ParameterError(
f"right input has unsupported shape {right.shape} for downmix=False"
)
if norm and not onesided:
# Only normalize here if we had both channels on input
output /= 2
return output
[docs]
def to_multi(
*signals: np.ndarray,
downmix: bool = True,
pad: bool = True,
norm: bool = True,
out: np.ndarray | None = None,
) -> np.ndarray:
"""Combine multiple signals into a multi-channel signal.
Parameters
----------
*signals : np.ndarray [shape=(..., n)]
One or more signals to combine into a multi-channel signal.
Each input signal can be mono or multi-channel, and they need not all have
identical lengths.
downmix : bool
If `True`, downmix each input signal to mono before combining.
If `False`, the input signals are additively combined, and all must have
identical channel layouts (shape excluding the trailing length dimension).
pad : bool
If `True`, pad the output to match the length of the longest input signal.
If `False`, trim the output to match the length of the shortest input signal.
norm : bool
If `True` (default), signals are combined by averaging.
If `False`, signals are combined by summing.
out : np.ndarray [shape=(m, n_out) or shape=(..., n_out)] or None
A pre-allocated output buffer. If provided, it must match the result
shape, and its dtype must be at least as precise as the result dtype;
it is zeroed and filled in place and returned, avoiding a new
allocation. If ``None`` (default), a new array is allocated.
Returns
-------
y_multi : np.ndarray [shape=(m, n_out) or shape=(..., n_out)]
Multi-channel combination of the input signals, where `m` corresponds
to the number of signals (if `downmix=True`), and `n_out` is the length
of the output signal determined by the padding mode.
If ``out`` was provided, this is the same object as ``out``.
See Also
--------
to_mono
to_stereo
util.fix_length
Examples
--------
Combine three mono signals of different lengths into a multi-channel signal
>>> y1 = librosa.tone(440.0, sr=22050, duration=1.0)
>>> y2 = librosa.tone(550.0, sr=22050, duration=1.0)
>>> y3 = librosa.tone(660.0, sr=22050, duration=2.0)
>>> y_multi = librosa.to_multi(y1, y2, y3, pad=True)
>>> y_multi.shape
(3, 44100)
"""
# Validate the buffer(s) and identify lengths, dtypes, etc
if not signals:
raise ParameterError("At least one signal must be provided.")
n_min, n_max = signals[0].shape[-1], signals[0].shape[-1]
dtype = signals[0].dtype
if downmix:
# If all channels are downmixed to mono,
# then the layout will be (n_signals, n_out)
channel_layout = tuple((len(signals),))
else:
# If we're not downmixing, then all have to have
# identical channel layout
channel_layout = tuple(signals[0].shape[:-1])
for y in signals:
util.valid_audio(y)
n_min = min(n_min, y.shape[-1])
n_max = max(n_max, y.shape[-1])
dtype = np.promote_types(dtype, y.dtype)
if not downmix and y.shape[:-1] != channel_layout:
raise ParameterError(
f"Cannot combine signals with different channel layouts {y.shape[:-1]} when downmix=False"
)
# Allocate and initialize the output buffer
if pad:
size = n_max
else:
size = n_min
output = _init_output(out, (*channel_layout, size), dtype)
if downmix:
# Downmix each signal to mono and mix into the output
for i, y in enumerate(signals):
y = y.astype(dtype, copy=False)
if y.shape[-1] == size:
# No length adjustment needed: downmix straight into the row
to_mono(y, norm=norm, out=output[i])
else:
# Length differs, so fix_length must allocate; a pass-through
# of out= is not possible here
output[i] = util.fix_length(to_mono(y, norm=norm), size=size, axis=-1)
else:
# Truncate and mix into the output
for y in signals:
output += util.fix_length(y, size=size, axis=-1)
if norm:
# Divide by the number of input signals given
output /= len(signals)
return output
[docs]
@cache(level=20)
def resample(
y: np.ndarray,
*,
orig_sr: float,
target_sr: float,
res_type: str = "soxr_hq",
fix: bool = True,
scale: bool = False,
axis: int = -1,
**kwargs: Any,
) -> np.ndarray:
"""Resample a time series from orig_sr to target_sr
By default, this uses a high-quality method (`soxr_hq`) 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, with `n` samples along the specified axis.
orig_sr : number > 0 [scalar]
original sampling rate of ``y``
target_sr : number > 0 [scalar]
target sampling rate
res_type : str (default: `soxr_hq`)
resample type
'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`.
'soxr_qq'
`soxr` Quick cubic interpolation (very fast, but not bandlimited)
'kaiser_best'
`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, but not bandlimited)
'zero_order_hold'
`samplerate` repeat the last value between samples. (very fast, but not bandlimited)
'sinc_best', 'sinc_medium' or 'sinc_fastest'
`samplerate` high-, medium-, and low-quality bandlimited sinc interpolation.
.. note::
Not all options yield a bandlimited interpolator. If you use `soxr_qq`, `polyphase`,
`linear`, or `zero_order_hold`, you need to be aware of possible aliasing effects.
.. note::
`samplerate` and `resampy` are not installed with `librosa`.
To use `samplerate` or `resampy`, they should be installed manually::
$ pip install samplerate
$ pip install resampy
.. 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.
axis : int
The target axis along which to resample. Defaults to the trailing axis.
**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`` along the target axis
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.loadx('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)
if orig_sr == target_sr:
return y
ratio = float(target_sr) / orig_sr
n_samples = int(np.ceil(y.shape[axis] * ratio))
if res_type in ("scipy", "fft"):
import scipy.signal
y_hat = scipy.signal.resample(y, n_samples, axis=axis)
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."
)
import scipy.signal
# 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=axis
)
elif res_type in (
"linear",
"zero_order_hold",
"sinc_best",
"sinc_fastest",
"sinc_medium",
):
# Use numpy to vectorize the resampler along the target axis
# This is because samplerate does not support ndim>2 generally.
y_hat = np.apply_along_axis(
samplerate.resample, axis=axis, arr=y, ratio=ratio, converter_type=res_type
)
elif res_type.startswith("soxr"):
# Use numpy to vectorize the resampler along the target axis
# This is because soxr does not support ndim>2 generally.
y_hat = np.apply_along_axis(
soxr.resample,
axis=axis,
arr=y,
in_rate=orig_sr,
out_rate=target_sr,
quality=res_type,
)
else:
y_hat = resampy.resample(y, orig_sr, target_sr, filter=res_type, axis=axis)
if fix:
y_hat = util.fix_length(y_hat, size=n_samples, axis=axis, **kwargs)
if scale:
y_hat /= np.sqrt(ratio)
# Match dtypes
return np.asarray(y_hat, dtype=y.dtype)
[docs]
def get_duration(
*,
y: np.ndarray | None = None,
sr: float = 22050,
S: np.ndarray | None = None,
n_fft: int = 2048,
hop_length: int = 512,
center: bool = True,
path: str | os.PathLike[Any] | None = None,
) -> float:
"""Compute the duration (in seconds) of an audio time series, feature matrix, or filename.
Examples
--------
>>> # Load an example audio file
>>> y, sr = librosa.loadx('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.loadx('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]
length of the FFT frame used when computing ``S``
hop_length : int > 0 [ scalar]
number of audio samples between columns of ``S``
center : bool
- If ``True``, ``S[:, t]`` is centered at ``y[t * hop_length]``
- If ``False``, then ``S[:, t]`` begins at ``y[t * hop_length]``
path : str, path, or file-like
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 ``path`` are provided.
Notes
-----
`get_duration` can be applied to a file (``path``), 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., ``path`` and ``S``),
then ``path`` takes precedence over ``S``, and ``S`` takes precedence over
``(y, sr)``.
"""
if path is not None:
return sf.info(path).duration # type: ignore
if y is None:
if S is None:
raise ParameterError("At least one of (y, sr), S, or path 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: str | int | sf.SoundFile | BinaryIO) -> float:
"""Get the sampling rate for a given file.
Parameters
----------
path : str, 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
"""
if isinstance(path, sf.SoundFile):
return path.samplerate # type: ignore
return sf.info(path).samplerate # type: ignore
[docs]
@cache(level=20)
def autocorrelate(
y: np.ndarray, *, max_size: int | None = None, axis: int = -1
) -> np.ndarray:
"""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.loadx('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]))
real = not np.iscomplexobj(y)
# Pad out the signal to support full-length auto-correlation
n_pad = scipy.fft.next_fast_len(2 * y.shape[axis] - 1, real=real)
autocorr: np.ndarray
if real:
# Compute the power spectrum along the chosen axis
powspec = util.abs2(scipy.fft.rfft(y, n=n_pad, axis=axis))
# Convert back to time domain
autocorr = scipy.fft.irfft(powspec, n=n_pad, axis=axis)
else:
# Compute the power spectrum along the chosen axis
powspec = util.abs2(scipy.fft.fft(y, n=n_pad, axis=axis))
# Convert back to time domain
autocorr = scipy.fft.ifft(powspec, n=n_pad, axis=axis)
# Slice down to max_size
subslice = [slice(None)] * autocorr.ndim
subslice[axis] = slice(max_size)
autocorr_slice: np.ndarray = autocorr[tuple(subslice)]
return autocorr_slice
[docs]
def lpc(y: np.ndarray, *, order: int, axis: int = -1) -> np.ndarray:
"""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 Acoustics, 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.loadx('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.loadx('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 util.is_positive_int(order):
raise ParameterError(f"order={order} must be an integer > 0")
util.valid_audio(y)
# 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 np.swapaxes(
__lpc(y, order, ar_coeffs, ar_coeffs_prev, reflect_coeff, den, epsilon), 0, axis
)
@jit(nopython=True, cache=True) # type: ignore
def __lpc(
y: np.ndarray,
order: int,
ar_coeffs: np.ndarray,
ar_coeffs_prev: np.ndarray,
reflect_coeff: np.ndarray,
den: np.ndarray,
epsilon: float,
) -> np.ndarray:
# 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 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
@stencil # type: ignore
def _zc_stencil(x: np.ndarray, threshold: float, zero_pos: bool) -> NDArray[np.bool]:
"""Stencil to compute zero crossings"""
x0 = x[0]
if -threshold <= x0 <= threshold:
x0 = 0
x1 = x[-1]
if -threshold <= x1 <= threshold:
x1 = 0
if zero_pos:
return np.signbit(x0) != np.signbit(x1) # type: ignore
else:
return np.sign(x0) != np.sign(x1) # type: ignore
@guvectorize(
"(n),(),()->(n)",
cache=True,
nopython=True,
) # type: ignore
def _zc_wrapper(
x: np.ndarray,
threshold: float,
zero_pos: bool,
y: NDArray[np.bool],
) -> None: # pragma: no cover
"""Vectorized wrapper for zero crossing stencil"""
y[:] = _zc_stencil(x, threshold, zero_pos)
[docs]
@cache(level=20)
def zero_crossings(
y: np.ndarray,
*,
threshold: float = 1e-10,
ref_magnitude: float | Callable | None = None,
pad: bool = True,
zero_pos: bool = True,
axis: int = -1,
) -> NDArray[np.bool]:
"""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
If non-zero, 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 : bool
If ``True``, then ``y[0]`` is considered a valid zero-crossing.
zero_pos : bool
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=bool]
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]),)
"""
if callable(ref_magnitude):
threshold = threshold * ref_magnitude(np.abs(y))
elif ref_magnitude is not None:
threshold = threshold * ref_magnitude
yi = y.swapaxes(-1, axis)
z = np.empty_like(y, dtype=bool)
zi = z.swapaxes(-1, axis)
_zc_wrapper(yi, threshold, zero_pos, zi)
zi[..., 0] = pad
return z
[docs]
def clicks(
*,
times: _SequenceLike[_FloatLike_co] | None = None,
frames: _SequenceLike[_IntLike_co] | None = None,
sr: float = 22050,
hop_length: int = 512,
click_freq: float = 1000.0,
click_duration: float = 0.1,
click: np.ndarray | None = None,
length: int | None = None,
) -> NDArray[np.float32]:
"""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.loadx('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(S, vscale='dBFS[power]',
... 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
positions: np.ndarray
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)
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(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]
def tone(
frequency: _FloatLike_co,
*,
sr: float = 22050,
length: int | None = None,
duration: float | None = None,
phi: float | None = None,
) -> _Array1D[np.float64]:
"""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, sr=22050)
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(S, vscale='dBFS[power]',
... 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 = int(duration * sr)
if phi is None:
phi = -np.pi * 0.5
y: np.ndarray = np.cos(2 * np.pi * frequency * np.arange(length) / sr + phi)
return y
[docs]
def chirp(
*,
fmin: _FloatLike_co,
fmax: _FloatLike_co,
sr: float = 22050,
length: int | None = None,
duration: float | None = None,
linear: bool = False,
phi: float | None = None,
) -> _Array1D[np.float64]:
"""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 : bool
- 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, sr=22050)
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, sr=22050)
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(S_exponential, vscale='dBFS',
... 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(S_linear, vscale='dBFS',
... 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"
import scipy.signal
y: np.ndarray = scipy.signal.chirp( # type: ignore[call-overload]
np.arange(int(duration * sr)) / sr,
fmin,
duration,
fmax,
method=method,
phi=phi / np.pi * 180, # scipy.signal.chirp uses degrees for phase offset
)
return y
[docs]
def mu_compress(
x: np.ndarray | _FloatLike_co, *, mu: float = 255, quantize: bool = True
) -> np.ndarray:
"""Compression by μ-law
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(
f"mu-law compression parameter mu={mu} must be strictly positive."
)
if np.any(x < -1) or np.any(x > 1):
raise ParameterError(f"mu-law input x={x} must be in the range [-1, +1].")
x_comp: np.ndarray = np.sign(x) * np.log1p(mu * np.abs(x)) / np.log1p(mu)
if quantize:
y: np.ndarray = (
np.digitize(
x_comp, np.linspace(-1, 1, num=int(1 + mu), endpoint=True), right=True
)
- int(mu + 1) // 2
)
return y
return x_comp
@overload
def mu_expand(x: _FloatLike_co, *, mu: float = 255.0, quantize: bool = True) -> np.floating: ...
@overload
def mu_expand(x: np.ndarray, *, mu: float = 255.0, quantize: bool = True) -> np.ndarray: ...
[docs]
def mu_expand(
x: np.ndarray | _FloatLike_co, *, mu: float = 255.0, quantize: bool = True
) -> np.ndarray | np.floating:
"""Expansion by μ-law
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 : bool
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(
f"Inverse mu-law compression parameter mu={mu} must be strictly positive."
)
if quantize:
x = x * 2.0 / (1 + mu)
if np.any(x < -1) or np.any(x > 1):
raise ParameterError(
f"Inverse mu-law input x={x} must be in the range [-1, +1]."
)
y: np.ndarray | np.floating = np.sign(x) / mu * (np.power(1 + mu, np.abs(x)) - 1)
return y