def pulse_finder_file(file,
freq_range,
pulse_rate_range,
window_len,
rejection_bands=None,
plot=False):
"""a wrapper for pulse_finder with takes an audio file path as an argument
creates the audio object and spectrogram within the function
Args:
file: path to an audio file
freq_range: range to bandpass the spectrogram, in Hz
pulse_rate_range: how many pulses per second? (where to look in the fft of the smoothed-amplitude), in Hz
rejection_bands: list of frequency bands to subtract from the desired freq_range
plot=False : if True, plot figures
Returns:
array of pulse_score: pulse score (float) for each time window
array of time: start time of each window
"""
# make spectrogram from file path
audio = Audio(file)
spec = Spectrogram.from_audio(audio)
pulse_scores, window_start_times = pulse_finder(spec, freq_range,
pulse_rate_range,
window_len,
rejection_bands, plot)
return pulse_scores, window_start_times
def image_from_audio(self, audio, mode="RGB"):
""" Create a PIL image from audio
Inputs:
audio: audio object
mode: PIL image mode, e.g. "L" or "RGB" [default: RGB]
"""
spectrogram = Spectrogram.from_audio(audio)
return spectrogram.to_image(shape=(self.width, self.height), mode=mode)
def test_spectrogram_shape_of_veryshort(veryshort_wav_str):
audio = Audio.from_file(veryshort_wav_str, sample_rate=22050)
spec = Spectrogram.from_audio(audio, overlap_samples=384)
assert spec.spectrogram.shape == (257, 21)
assert spec.frequencies.shape == (257, )
assert spec.times.shape == (21, )
assert isclose(spec.window_length(), 0.02321995465, abs_tol=1e-4)
assert isclose(spec.window_step(), 0.005804988662, abs_tol=1e-4)
assert isclose(spec.duration(), audio.duration(), abs_tol=1e-2)
assert isclose(spec.window_start_times()[0], 0, abs_tol=1e-4)
def test_ribbit():
path = "./tests/audio/silence_10s.mp3"
audio = Audio.from_file(path, sample_rate=22050)
spec = Spectrogram.from_audio(audio)
scores, times = ribbit.ribbit(
spec,
pulse_rate_range=[5, 10],
signal_band=[1000, 2000],
window_len=5.0,
noise_bands=[[0, 200]],
plot=True,
)
assert len(scores) > 0
def test_pulse_finder():
path = "./tests/silence_10s.mp3"
audio = Audio(path)
spec = Spectrogram.from_audio(audio)
scores, times = pulse_finder(
spec,
pulse_rate_range=[5, 10],
freq_range=[1000, 2000],
window_len=5.0,
rejection_bands=[[0, 200]],
plot=True,
)
assert len(scores) > 0
def process_audio(func_name, audio_io):
"""
Check inputs and return spectrogram images
Check that inputs are in correct format (single-channel,
between 5-20 seconds), and convert to images
Args:
func_name (str): name of calling function (for printing)
audio_io (bytes):
Returns:
array of 299x299 images, each representing up to
5s of the original audio
"""
print(f"runserver.py: {func_name}() checking inputs")
# Just return error if no data was posted
if not audio_io:
return {"error": "No data was given with post?"}
# Make sure we can load the data given to us
print(f"runserver.py: {func_name}() loading samples")
try:
audio = Audio.from_bytesio(
audio_io, sample_rate=22050, resample_type="kaiser_fast"
)
except:
return {"error": "Unable to load audio, multi-chennel input is ignored"}
print(
f"runserver.py: {func_name}() loaded samples at sample_rate {audio.sample_rate}"
)
# Check the duration is between 5 and 20 seconds
duration = audio.duration()
if duration < 5:
return {"error": "Audio is shorter than 5 seconds"}
elif duration > 20:
return {"error": "Audio is longer than 20 seconds"}
# 1. split audio into 5 second chunks
# 2. generate spectrograms
# 3. generate images
audio_splits = birds_detector.split_audio(audio)
spectrograms = [Spectrogram.from_audio(x) for x in audio_splits]
images = [x.to_image(shape=(224, 224)) for x in spectrograms]
print(f"runserver.py: {func_name}(), opening audio as spectrograms")
return {"images": images}
def test_summarize_top_scores(gpt_path):
df = pd.DataFrame(columns=[
"species",
"pulse_rate_low",
"pulse_rate_high",
"low_f",
"high_f",
"reject_low",
"reject_high",
"window_length",
])
df.at[0, :] = ["sp1", 5, 10, 1000, 2000, 0, 500, 1.0]
df.at[1, :] = ["sp2", 10, 15, 1000, 2000, 0, 500, 1.0]
audio = Audio.from_file(gpt_path, sample_rate=32000)
spec = Spectrogram.from_audio(audio, overlap_samples=256)
df = ribbit.pulse_finder_species_set(spec, df)
ribbit.summarize_top_scores(["1", "2"], [df, df], scale_factor=10.0)
def test_ribbit_short_audio(veryshort_wav_str):
audio = Audio.from_file(veryshort_wav_str, sample_rate=22050)
spec = Spectrogram.from_audio(audio,
window_samples=512,
overlap_samples=256,
decibel_limits=(-100, -20))
df = ribbit.ribbit(
spec,
pulse_rate_range=[5, 10],
signal_band=[1000, 2000],
clip_duration=5.0,
clip_overlap=2.5,
final_clip=None,
noise_bands=[[0, 200]],
plot=False,
)
assert len(df) == 0
def test_pulsefinder_species_set(gpt_path):
df = pd.DataFrame(columns=[
"species",
"pulse_rate_low",
"pulse_rate_high",
"low_f",
"high_f",
"reject_low",
"reject_high",
"window_length",
])
df.at[0, :] = ["sp1", 5, 10, 1000, 2000, 0, 500, 1.0]
df.at[1, :] = ["sp2", 10, 15, 1000, 2000, 0, 500, 1.0]
audio = Audio.from_file(gpt_path, sample_rate=32000)
spec = Spectrogram.from_audio(audio, overlap_samples=256)
df = ribbit.pulse_finder_species_set(spec, df)
assert type(df) == pd.DataFrame
def test_ribbit_high_spec_overlap(gpt_path):
"""spec params should not effect number of clips in results"""
audio = Audio.from_file(gpt_path, sample_rate=22050).trim(0, 16)
spec = Spectrogram.from_audio(audio,
window_samples=512,
overlap_samples=500,
decibel_limits=(-100, -20))
df = ribbit.ribbit(
spec,
pulse_rate_range=[5, 10],
signal_band=[1000, 2000],
clip_duration=5.0,
clip_overlap=0,
final_clip=None,
noise_bands=[[0, 200]],
plot=False,
)
assert len(df) == 3
assert isclose(max(df["start_time"]), 10.0, abs_tol=1e-4)
def test_ribbit(gpt_path):
audio = Audio.from_file(gpt_path, sample_rate=22050).trim(0, 16)
spec = Spectrogram.from_audio(audio,
window_samples=512,
overlap_samples=256,
decibel_limits=(-100, -20))
df = ribbit.ribbit(
spec,
pulse_rate_range=[5, 10],
signal_band=[1000, 2000],
clip_duration=5.0,
clip_overlap=0,
final_clip=None,
noise_bands=[[0, 200]],
plot=False,
)
assert len(df) == 3
assert isclose(max(df["score"]), 0.0392323, abs_tol=1e-4)
def test_pulsefinder_species_set():
path = "./tests/great_plains_toad.wav"
df = pd.DataFrame(columns=[
"species",
"pulse_rate_low",
"pulse_rate_high",
"low_f",
"high_f",
"reject_low",
"reject_high",
"window_length",
])
df.at[0, :] = ["sp1", 5, 10, 1000, 2000, 0, 500, 1.0]
df.at[1, :] = ["sp2", 10, 15, 1000, 2000, 0, 500, 1.0]
audio = Audio(path, sample_rate=32000)
spec = Spectrogram.from_audio(audio, overlap_samples=256)
df = pulse_finder_species_set(spec, df)
assert type(df) == pd.DataFrame
def test_spectrogram_shape_of_veryshort(veryshort_wav_str):
audio = Audio.from_file(veryshort_wav_str, sample_rate=22050)
spec = Spectrogram.from_audio(audio, overlap_samples=384)
assert spec.spectrogram.shape == (257, 21)
assert spec.frequencies.shape == (257, )
assert spec.times.shape == (21, )
def test_spectrogram_raises_typeerror():
with pytest.raises(TypeError):
Spectrogram.from_audio("not samples")
def go(self, audio):
return Spectrogram.from_audio(audio, **self.params)