Download this notebook by clicking on the download icon logo or find it in the repository

PacificSoundClassifyBlueA

Distributed under the terms of the GPL License
Maintainer: dcline@mbari.org
Authors: Danelle Cline dcline@mbari.org, John Ryan ryjo@mbari.org

Kernel Selection¶

If running in SageMaker, the Python 3 (Data Science) kernel is sufficient. The Python 3 (TensorFlow 2.3 Python 3.7 GPU Optimized) will run the inference code faster for a higher cost. For more advanced users, SageMaker Batch Transform is recommended to process data in bulk.

Applying Machine Learning to classify blue whale A calls¶

Essential to detection and classification of marine mammal vocalizations are the distinct acoustic attributes of those vocalizations. Machine learning (ML) is an effective way to recognize acoustic attributes and reliably classify such vocalizations.

In this brief tutorial, we will:

tap into an extensive (6+ years and growing) archive of sound recordings from a deep-sea location along the eastern margin of the North Pacific Ocean,
illustrate the beautiful songs produced by baleen whales, and
demonstrate the application of ML to classify one of the three types of calls produced by blue whales in their songs.

If you use this data set or tutorial, please cite our project[1].

Install dependencies¶

First, let's install dependencies and include all packages used in this tutorial. This only needs to be done once for the duration of this notebook.

Sound library install¶

If running on Linux, you may need to install the system library libsndfile. On Windows and OS X, pip install soundfile will install it for you.

In [1]:

# import platform
# if 'Linux' in platform.uname():
#     !apt-get update -y && apt-get install -y libsndfile1
# import platform
# if 'Linux' in platform.uname():
#     !apt-get update -y && apt-get install -y libsndfile1

In [2]:

!pip install tensorflow==2.4.1 --quiet
!pip install boto3 --quiet
!pip install oceansoundscape==1.0.6  --quiet
!pip install tensorflow==2.4.1 --quiet
!pip install boto3 --quiet
!pip install oceansoundscape==1.0.6  --quiet

WARNING: You are using pip version 21.2.4; however, version 21.3.1 is available.
You should consider upgrading via the '/raid/dcline-admin/docs/.venv/bin/python3.7 -m pip install --upgrade pip' command.
WARNING: You are using pip version 21.2.4; however, version 21.3.1 is available.
You should consider upgrading via the '/raid/dcline-admin/docs/.venv/bin/python3.7 -m pip install --upgrade pip' command.
WARNING: You are using pip version 21.2.4; however, version 21.3.1 is available.
You should consider upgrading via the '/raid/dcline-admin/docs/.venv/bin/python3.7 -m pip install --upgrade pip' command.

In [3]:

import boto3
from botocore import UNSIGNED
from botocore.client import Config
import cv2
from oceansoundscape.spectrogram.signal import psd_1sec
from oceansoundscape.raven import BLEDParser
from oceansoundscape.spectrogram import conf, denoise, utils
import os
from pathlib import Path
import soundfile as sf
import json
import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
from matplotlib.patches import Rectangle
import boto3
from botocore import UNSIGNED
from botocore.client import Config
import cv2
from oceansoundscape.spectrogram.signal import psd_1sec
from oceansoundscape.raven import BLEDParser
from oceansoundscape.spectrogram import conf, denoise, utils
import os
from pathlib import Path
import soundfile as sf
import json
import numpy as np
import matplotlib.pyplot as plt
import tensorflow as tf
from matplotlib.patches import Rectangle

2021-11-22 16:29:22.683973: W tensorflow/stream_executor/platform/default/dso_loader.cc:60] Could not load dynamic library 'libcudart.so.11.0'; dlerror: libcudart.so.11.0: cannot open shared object file: No such file or directory
2021-11-22 16:29:22.683999: I tensorflow/stream_executor/cuda/cudart_stub.cc:29] Ignore above cudart dlerror if you do not have a GPU set up on your machine.

Why use 2 kHz data?¶

Because we are studying the low-frequency calls of blue whales, we don't need the original recordings sampled at 256 kHz. Instead, we will use the 2 kHz decimated data in WAV format; these are stored in an s3 bucket named pacific-sound-2khz. For more information on the storage and organization of the 2kHz data, please see the 2kHz example.

List the contents of a monthly directory¶

Between August 2015 and July 2021 (a 6-year period), the highest levels of blue whale song activity off central California were detected during November 2017. Let's start by listing the daily 2 kHz files for that month.

In [4]:

s3 = boto3.client('s3',
    aws_access_key_id='',
    aws_secret_access_key='',
    config=Config(signature_version=UNSIGNED))

year = "2017"
month = "11"
bucket = 'pacific-sound-2khz'

for obj in s3.list_objects_v2(Bucket=bucket, Prefix=f'{year}/{month}')['Contents']:
    print(obj['Key'])
s3 = boto3.client('s3',
    aws_access_key_id='',
    aws_secret_access_key='',
    config=Config(signature_version=UNSIGNED))

year = "2017"
month = "11"
bucket = 'pacific-sound-2khz'

for obj in s3.list_objects_v2(Bucket=bucket, Prefix=f'{year}/{month}')['Contents']:
    print(obj['Key'])

2017/11/MARS-20171101T000000Z-2kHz.wav
2017/11/MARS-20171102T000000Z-2kHz.wav
2017/11/MARS-20171103T000000Z-2kHz.wav
2017/11/MARS-20171104T000000Z-2kHz.wav
2017/11/MARS-20171105T000000Z-2kHz.wav
2017/11/MARS-20171106T000000Z-2kHz.wav
2017/11/MARS-20171107T000000Z-2kHz.wav
2017/11/MARS-20171108T000000Z-2kHz.wav
2017/11/MARS-20171109T000000Z-2kHz.wav
2017/11/MARS-20171110T000000Z-2kHz.wav
2017/11/MARS-20171111T000000Z-2kHz.wav
2017/11/MARS-20171112T000000Z-2kHz.wav
2017/11/MARS-20171113T000000Z-2kHz.wav
2017/11/MARS-20171114T000000Z-2kHz.wav
2017/11/MARS-20171115T000000Z-2kHz.wav
2017/11/MARS-20171116T000000Z-2kHz.wav
2017/11/MARS-20171117T000000Z-2kHz.wav
2017/11/MARS-20171118T000000Z-2kHz.wav
2017/11/MARS-20171119T000000Z-2kHz.wav
2017/11/MARS-20171120T000000Z-2kHz.wav
2017/11/MARS-20171121T000000Z-2kHz.wav
2017/11/MARS-20171122T000000Z-2kHz.wav
2017/11/MARS-20171123T000000Z-2kHz.wav
2017/11/MARS-20171124T000000Z-2kHz.wav
2017/11/MARS-20171125T000000Z-2kHz.wav
2017/11/MARS-20171126T000000Z-2kHz.wav
2017/11/MARS-20171127T000000Z-2kHz.wav
2017/11/MARS-20171128T000000Z-2kHz.wav
2017/11/MARS-20171129T000000Z-2kHz.wav
2017/11/MARS-20171130T000000Z-2kHz.wav
2017/11/copy.sh

A view of baleen whale song¶

Let's produce a spectrogram with sufficient resolution in time and frequency to see the blue whale song with enough resolution to visually identify. We'll limit the exercise to a single hour from a day with calls of variable received intensity (signal strength).

Download a single 2 kHz file¶

In [5]:

year = "2017"
month = "11"
wav_filename = 'MARS-20171101T000000Z-2kHz.wav'
bucket = 'pacific-sound-2khz'
key = f'{year}/{month}/{wav_filename}'

s3 = boto3.resource('s3',
    aws_access_key_id='',
    aws_secret_access_key='',
    config=Config(signature_version=UNSIGNED))
 
# only download if needed
if not Path(wav_filename).exists():

    # Alternatively, it can be downloaded directly in SageMaker with
    # !aws s3 cp s3://{bucket}/{key} . 

    print('Downloading') 
    s3.Bucket(bucket).download_file(key, wav_filename)
    print('Done')
year = "2017"
month = "11"
wav_filename = 'MARS-20171101T000000Z-2kHz.wav'
bucket = 'pacific-sound-2khz'
key = f'{year}/{month}/{wav_filename}'

s3 = boto3.resource('s3',
    aws_access_key_id='',
    aws_secret_access_key='',
    config=Config(signature_version=UNSIGNED))
 
# only download if needed
if not Path(wav_filename).exists():

    # Alternatively, it can be downloaded directly in SageMaker with
    # !aws s3 cp s3://{bucket}/{key} . 

    print('Downloading') 
    s3.Bucket(bucket).download_file(key, wav_filename)
    print('Done') 

Subset to the 5th hour of the day¶

In [6]:

sample_rate = int(2e3)
start_hour = 5
start_frame = int(sample_rate * start_hour * 3600)
duration_frames =  int(sample_rate* 3600)

pacsound_file = sf.SoundFile(wav_filename)
pacsound_file.seek(start_frame)
x = pacsound_file.read(duration_frames, dtype='float32')
sample_rate = int(2e3)
start_hour = 5
start_frame = int(sample_rate * start_hour * 3600)
duration_frames =  int(sample_rate* 3600)

pacsound_file = sf.SoundFile(wav_filename)
pacsound_file.seek(start_frame)
x = pacsound_file.read(duration_frames, dtype='float32')

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
/tmp/ipykernel_1886408/3365153408.py in <module>
      4 duration_frames =  int(sample_rate* 3600)
      5 
----> 6 pacsound_file = sf.SoundFile(wav_filename)
      7 pacsound_file.seek(start_frame)
      8 x = pacsound_file.read(duration_frames, dtype='float32')

/raid/dcline-admin/pacific-sound/docs/.venv/lib/python3.7/site-packages/soundfile.py in __init__(self, file, mode, samplerate, channels, subtype, endian, format, closefd)
    625         self._info = _create_info_struct(file, mode, samplerate, channels,
    626                                          format, subtype, endian)
--> 627         self._file = self._open(file, mode_int, closefd)
    628         if set(mode).issuperset('r+') and self.seekable():
    629             # Move write position to 0 (like in Python file objects)

/raid/dcline-admin/pacific-sound/docs/.venv/lib/python3.7/site-packages/soundfile.py in _open(self, file, mode_int, closefd)
   1180             raise TypeError("Invalid file: {0!r}".format(self.name))
   1181         _error_check(_snd.sf_error(file_ptr),
-> 1182                      "Error opening {0!r}: ".format(self.name))
   1183         if mode_int == _snd.SFM_WRITE:
   1184             # Due to a bug in libsndfile version <= 1.0.25, frames != 0

/raid/dcline-admin/pacific-sound/docs/.venv/lib/python3.7/site-packages/soundfile.py in _error_check(err, prefix)
   1353     if err != 0:
   1354         err_str = _snd.sf_error_number(err)
-> 1355         raise RuntimeError(prefix + _ffi.string(err_str).decode('utf-8', 'replace'))
   1356 
   1357 

RuntimeError: Error opening 'MARS-20171101T000000Z-2kHz.wav': File contains data in an unknown format.

Plot the full 2 kHz spectrogram¶

Lots of biophony (sounds of ocean life) are represented in this spectrogram. Humpback whale songs are dominant above ~ 100 Hz, while blue and fin whale songs are dominant below ~ 100 Hz. The energy of blue whale A calls is largely between ~70 and 90 Hz (between the white dashed lines).

For more details about creating a calibrated spectrogram, see the 2 kHz tutorial.

In [7]:

sg, f = psd_1sec(x, sample_rate, 177.9) # create calibrated psd
plt.figure(dpi=300)
plt.axhline(73,linestyle='--', color='white')
plt.axhline(91,linestyle='--', color='white')
plt.imshow(sg,extent=[0, 3600, min(f), max(f)],aspect='auto',origin='lower',vmin=30,vmax=100)
plt.yscale('log')
plt.ylim(10,1000)
plt.colorbar()
plt.xlabel('November 01, 2017 Hour 5')
plt.ylabel('Frequency (Hz)')
plt.title('Calibrated spectrum levels')
sg, f = psd_1sec(x, sample_rate, 177.9) # create calibrated psd
plt.figure(dpi=300)
plt.axhline(73,linestyle='--', color='white')
plt.axhline(91,linestyle='--', color='white')
plt.imshow(sg,extent=[0, 3600, min(f), max(f)],aspect='auto',origin='lower',vmin=30,vmax=100)
plt.yscale('log')
plt.ylim(10,1000)
plt.colorbar()
plt.xlabel('November 01, 2017 Hour 5')
plt.ylabel('Frequency (Hz)')
plt.title('Calibrated spectrum levels')

---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
/tmp/ipykernel_1886408/602911774.py in <module>
----> 1 sg, f = psd_1sec(x, sample_rate, 177.9) # create calibrated psd
      2 plt.figure(dpi=300)
      3 plt.axhline(73,linestyle='--', color='white')
      4 plt.axhline(91,linestyle='--', color='white')
      5 plt.imshow(sg,extent=[0, 3600, min(f), max(f)],aspect='auto',origin='lower',vmin=30,vmax=100)

NameError: name 'x' is not defined

Band Limited Energy Detections¶

In this example, we will focus on classification of calls. Here, detections for potential Blue whale A calls have been precomputed using Band-Limited-Energy-Detectors BLEDs in the Raven software. There are many way to do detection, including spectrogram correlation and neural network approaches such as single-shot detectors to name a few. The best detector depends on the acoustic application and its cost/performance tradeoff.

The approach taken here is to use the BLED detectors to identify not only strong and clear calls, but also many uncertain signals, so we can avoid missing calls.

An output from the BLED detectors might look something like this following simple table.

Note that this only represents a subset of the detections in the 5th hour for the purposes of this simple example.

In [8]:

bled_file = 'MARS-20171101T000000Z-2kHz.wav.Table01.txt'
bled_file = 'MARS-20171101T000000Z-2kHz.wav.Table01.txt'

In [9]:

%%writefile {bled_file}
Selection	View	Channel	Low Freq (Hz)	High Freq (Hz)	Begin Time (s)	End Time (s)
Spectrogram 1	1	70.0	90.0	19076.535750	19088.027750
Spectrogram 1	1	70.0	90.0	19115.990750	19132.331750
Spectrogram 1	1	70.0	90.0	19145.565750	19161.217750
Spectrogram 1	1	70.0	90.0	19182.693750	19200.880750
Spectrogram 1	1	70.0	90.0	19221.550750	19235.590750
Spectrogram 1	1	70.0	90.0	19244.183750	19270.027750
Spectrogram 1	1	70.0	90.0	19310.171750	19328.345750
Spectrogram 1	1	70.0	90.0	19342.502750	19360.689750
Spectrogram 1	1	70.0	90.0	19373.676750	19389.107750
Spectrogram 1	1	70.0	90.0	19398.727750	19410.193750
Spectrogram 1	1	70.0	90.0	19421.425750	19436.804750
Spectrogram 1	1	70.0	90.0	19445.189750	19461.803750
Spectrogram 1	1	70.0	90.0	19494.706750	19507.173750
Spectrogram 1	1	70.0	90.0	19527.622750	19538.555750
Spectrogram 1	1	70.0	90.0	19548.071750	19561.370750
Spectrogram 1	1	70.0	90.0	19568.377750	19585.186750
Spectrogram 1	1	70.0	90.0	19625.057750	19645.753750
Spectrogram 1	1	70.0	90.0	19653.293750	19676.381750
Spectrogram 1	1	70.0	90.0	19680.944750	19704.734750
Spectrogram 1	1	70.0	90.0	19708.543750	19723.051750
Spectrogram 1	1	70.0	90.0	19760.907750	19778.990750
%%writefile {bled_file}
Selection	View	Channel	Low Freq (Hz)	High Freq (Hz)	Begin Time (s)	End Time (s)
Spectrogram 1	1	70.0	90.0	19076.535750	19088.027750
Spectrogram 1	1	70.0	90.0	19115.990750	19132.331750
Spectrogram 1	1	70.0	90.0	19145.565750	19161.217750
Spectrogram 1	1	70.0	90.0	19182.693750	19200.880750
Spectrogram 1	1	70.0	90.0	19221.550750	19235.590750
Spectrogram 1	1	70.0	90.0	19244.183750	19270.027750
Spectrogram 1	1	70.0	90.0	19310.171750	19328.345750
Spectrogram 1	1	70.0	90.0	19342.502750	19360.689750
Spectrogram 1	1	70.0	90.0	19373.676750	19389.107750
Spectrogram 1	1	70.0	90.0	19398.727750	19410.193750
Spectrogram 1	1	70.0	90.0	19421.425750	19436.804750
Spectrogram 1	1	70.0	90.0	19445.189750	19461.803750
Spectrogram 1	1	70.0	90.0	19494.706750	19507.173750
Spectrogram 1	1	70.0	90.0	19527.622750	19538.555750
Spectrogram 1	1	70.0	90.0	19548.071750	19561.370750
Spectrogram 1	1	70.0	90.0	19568.377750	19585.186750
Spectrogram 1	1	70.0	90.0	19625.057750	19645.753750
Spectrogram 1	1	70.0	90.0	19653.293750	19676.381750
Spectrogram 1	1	70.0	90.0	19680.944750	19704.734750
Spectrogram 1	1	70.0	90.0	19708.543750	19723.051750
Spectrogram 1	1	70.0	90.0	19760.907750	19778.990750

Overwriting MARS-20171101T000000Z-2kHz.wav.Table01.txt

Focus on blue whale A calls¶

Zooming in on a more narrow frequency and time, we can see the pulse-train nature of the A calls, an important distinction for classification. In this brief segment, we can see A calls of variable received intensity.

Here, we parse the detections and display the identifying Raven Selection number for reference.

Notes:

The first A call in this time period, which was quite strong, does not have a start marker because its start precedes the beginning of this audio segment.
The full detail of the A calls is still not visible at the resolution presented above. The full detail will become clear when we apply the model and examine individual classified calls.

In [10]:

# an optimum configuration for the call spectrogram generation is defined in the oceanscoundscape package based on extensive
# hyper parameter sweeps. Let's grab those here
blue_a_conf = conf.CONF_DICT['blueA']

# BLEDParser needs a path object
bled_path = Path(bled_file)

# parse the detections
parser = BLEDParser(bled_path, blue_a_conf, max_samples=len(x), sampling_rate=sample_rate)

# plot the spectrogram
plt.figure(dpi=300)
plt.imshow(sg,aspect='auto',origin='lower',vmin=30,vmax=100)

# add boxes for each normalized detection - detections are normalized to a fixed size (frequency/time) for classification
start_secs = start_hour * 3600
high_freq = blue_a_conf['high_freq']
low_freq = blue_a_conf['low_freq']
duration_secs = blue_a_conf['duration_secs']
freq_window = int(high_freq - low_freq)
for row, item in sorted(parser.data.iterrows()):
    detection_start = item.call_start/sample_rate - start_secs # detections are stored in sample numbers so here we convert back to time
    plt.gca().add_patch(Rectangle((detection_start,low_freq),duration_secs,freq_window,
                        edgecolor='black',
                        facecolor='none',
                        ls='--',
                        lw=1))
    plt.text(detection_start,high_freq - 8,item.Selection, size='x-small', rotation=45)

plt.yscale('linear')
plt.ylim(10,high_freq + 10)
plt.xlim(1000, 1800)
plt.xlabel('Second of 01 November 2017')
plt.ylabel('Frequency (Hz)')
plt.title('Detected blue whale A calls')
# an optimum configuration for the call spectrogram generation is defined in the oceanscoundscape package based on extensive
# hyper parameter sweeps. Let's grab those here
blue_a_conf = conf.CONF_DICT['blueA']

# BLEDParser needs a path object
bled_path = Path(bled_file)

# parse the detections
parser = BLEDParser(bled_path, blue_a_conf, max_samples=len(x), sampling_rate=sample_rate)

# plot the spectrogram
plt.figure(dpi=300)
plt.imshow(sg,aspect='auto',origin='lower',vmin=30,vmax=100)

# add boxes for each normalized detection - detections are normalized to a fixed size (frequency/time) for classification
start_secs = start_hour * 3600
high_freq = blue_a_conf['high_freq']
low_freq = blue_a_conf['low_freq']
duration_secs = blue_a_conf['duration_secs']
freq_window = int(high_freq - low_freq)
for row, item in sorted(parser.data.iterrows()):
    detection_start = item.call_start/sample_rate - start_secs # detections are stored in sample numbers so here we convert back to time
    plt.gca().add_patch(Rectangle((detection_start,low_freq),duration_secs,freq_window,
                        edgecolor='black',
                        facecolor='none',
                        ls='--',
                        lw=1))
    plt.text(detection_start,high_freq - 8,item.Selection, size='x-small', rotation=45)

plt.yscale('linear')
plt.ylim(10,high_freq + 10)
plt.xlim(1000, 1800)
plt.xlabel('Second of 01 November 2017')
plt.ylabel('Frequency (Hz)')
plt.title('Detected blue whale A calls')

---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
/tmp/ipykernel_1886408/2688997198.py in <module>
      7 
      8 # parse the detections
----> 9 parser = BLEDParser(bled_path, blue_a_conf, max_samples=len(x), sampling_rate=sample_rate)
     10 
     11 # plot the spectrogram

NameError: name 'x' is not defined

Classification of blue whale A calls¶

We can turn these detected audio signals into images, one for each detected signal, and classify them using a model trained for only A calls.

Create optimized spectrograms¶

Because the model we are demonstrating is image based, the first step to classify the calls is to create optimized spectrograms for each candidate detection.

In [11]:

# read the wav file
xc, sample_rate = sf.read(wav_filename)
 
# parse the detections, but this time let's use the entire wav file to allow for padding during denoising
parser_full = BLEDParser(bled_path, blue_a_conf, len(xc), sample_rate)

# grab the spectrogram parameters
call_width = int(blue_a_conf['duration_secs'] * sample_rate)
num_fft = blue_a_conf['num_fft']
hop_length = round(num_fft * (1 - conf.OVERLAP))

# generate optimized spectrograms
print(f'Denoising {bled_file} num detections {parser_full.num_detections}')

# denoise them and store the results in a temporary hdf file cache
cache = denoise.PCENSmooth(Path(f'{bled_path}.hdf'), parser_full, blue_a_conf, xc, sample_rate)

for row, item in sorted(parser_full.data.iterrows()):
    try:
        start = int(call_width / hop_length)
        end = int(2 * call_width / hop_length)

        data = cache.get_data(item.Selection)

        # subset the call, leaving off the padding for PCEN
        subset_data = data[:, start:end]

        # save to a denoised color image
        utils.ImageUtils.colorizeDenoise(subset_data, Path(item.image_filename))
    except Exception as ex:
        print(ex)
        continue

print('Done')
# read the wav file
xc, sample_rate = sf.read(wav_filename)
 
# parse the detections, but this time let's use the entire wav file to allow for padding during denoising
parser_full = BLEDParser(bled_path, blue_a_conf, len(xc), sample_rate)

# grab the spectrogram parameters
call_width = int(blue_a_conf['duration_secs'] * sample_rate)
num_fft = blue_a_conf['num_fft']
hop_length = round(num_fft * (1 - conf.OVERLAP))

# generate optimized spectrograms
print(f'Denoising {bled_file} num detections {parser_full.num_detections}')

# denoise them and store the results in a temporary hdf file cache
cache = denoise.PCENSmooth(Path(f'{bled_path}.hdf'), parser_full, blue_a_conf, xc, sample_rate)

for row, item in sorted(parser_full.data.iterrows()):
    try:
        start = int(call_width / hop_length)
        end = int(2 * call_width / hop_length)

        data = cache.get_data(item.Selection)

        # subset the call, leaving off the padding for PCEN
        subset_data = data[:, start:end]

        # save to a denoised color image
        utils.ImageUtils.colorizeDenoise(subset_data, Path(item.image_filename))
    except Exception as ex:
        print(ex)
        continue

print('Done')

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
/tmp/ipykernel_1886408/1437253740.py in <module>
      1 # read the wav file
----> 2 xc, sample_rate = sf.read(wav_filename)
      3 
      4 # parse the detections, but this time let's use the entire wav file to allow for padding during denoising
      5 parser_full = BLEDParser(bled_path, blue_a_conf, len(xc), sample_rate)

/raid/dcline-admin/pacific-sound/docs/.venv/lib/python3.7/site-packages/soundfile.py in read(file, frames, start, stop, dtype, always_2d, fill_value, out, samplerate, channels, format, subtype, endian, closefd)
    255     """
    256     with SoundFile(file, 'r', samplerate, channels,
--> 257                    subtype, endian, format, closefd) as f:
    258         frames = f._prepare_read(start, stop, frames)
    259         data = f.read(frames, dtype, always_2d, fill_value, out)

/raid/dcline-admin/pacific-sound/docs/.venv/lib/python3.7/site-packages/soundfile.py in __init__(self, file, mode, samplerate, channels, subtype, endian, format, closefd)
    625         self._info = _create_info_struct(file, mode, samplerate, channels,
    626                                          format, subtype, endian)
--> 627         self._file = self._open(file, mode_int, closefd)
    628         if set(mode).issuperset('r+') and self.seekable():
    629             # Move write position to 0 (like in Python file objects)

/raid/dcline-admin/pacific-sound/docs/.venv/lib/python3.7/site-packages/soundfile.py in _open(self, file, mode_int, closefd)
   1180             raise TypeError("Invalid file: {0!r}".format(self.name))
   1181         _error_check(_snd.sf_error(file_ptr),
-> 1182                      "Error opening {0!r}: ".format(self.name))
   1183         if mode_int == _snd.SFM_WRITE:
   1184             # Due to a bug in libsndfile version <= 1.0.25, frames != 0

/raid/dcline-admin/pacific-sound/docs/.venv/lib/python3.7/site-packages/soundfile.py in _error_check(err, prefix)
   1353     if err != 0:
   1354         err_str = _snd.sf_error_number(err)
-> 1355         raise RuntimeError(prefix + _ffi.string(err_str).decode('utf-8', 'replace'))
   1356 
   1357 

RuntimeError: Error opening 'MARS-20171101T000000Z-2kHz.wav': File contains data in an unknown format.

Display the spectrogram images¶

In [12]:

images = list( sorted((Path.cwd()).glob('*.jpg')))

plt.figure(figsize=(14,14))
plt.rcParams['axes.titlelocation'] = 'center'
for i, image_path in enumerate(images):
    # the selection number of part of the filename name,
    # e.g. selection 3 20150925T072604.53528828.53552892.sel.03.ch01.spectrogram.jpg
    selection = image_path.name.split('.')[4]
    image_path = str(image_path)
    plt.subplot(5, 5, i + 1)
    img = plt.imread(image_path)
    plt.title(selection)
    plt.imshow(img)
    plt.axis('off')
images = list( sorted((Path.cwd()).glob('*.jpg')))

plt.figure(figsize=(14,14))
plt.rcParams['axes.titlelocation'] = 'center'
for i, image_path in enumerate(images):
    # the selection number of part of the filename name,
    # e.g. selection 3 20150925T072604.53528828.53552892.sel.03.ch01.spectrogram.jpg
    selection = image_path.name.split('.')[4]
    image_path = str(image_path)
    plt.subplot(5, 5, i + 1)
    img = plt.imread(image_path)
    plt.title(selection)
    plt.imshow(img)
    plt.axis('off')