def getSimilarSongSegments(audiofile, feature_type = 'rh'):
samplerate, samplewidth, wavedata = audiofile_read(audiofile)
# here we set the feature type for the analysis, but we need to put it in a list [.]
fext = [feature_type]
segment_features = rp_extract(wavedata,
samplerate,
extract_rp = ('rp' in fext), # extract Rhythm Patterns features
extract_ssd = ('ssd' in fext), # extract Statistical Spectrum Descriptor
extract_sh = ('sh' in fext), # extract Statistical Histograms
extract_tssd = ('tssd' in fext), # extract temporal Statistical Spectrum Descriptor
extract_rh = ('rh' in fext), # extract Rhythm Histogram features
extract_trh = ('trh' in fext), # extract temporal Rhythm Histogram features
extract_mvd = ('mvd' in fext), # extract Modulation Frequency Variance Descriptor
skip_leadin_fadeout=0,
step_width=1,
return_segment_features = True)
search_features = segment_features[feature_type]
# IMPORTANT: we use query_features as a GLOBAL VARIABLE here
# that means getSimilarSongs MUST ALWAYS BE CALLED BEFORE THIS FUNCTION
query_feature_vector = query_features[feature_type]
# Searching similar song segments
sim_song_search = NearestNeighbors(n_neighbors = 1, metric='euclidean')
# TODO proper Scaling (Normalization) (see above)
sim_song_search.fit(search_features)
# Get the most similar song SEGMENTS indices
most_similar_segment_index = sim_song_search.kneighbors(query_feature_vector, return_distance=False)
# here you get the segment's sample position
# segment_features['segpos']
# and time positions in seconds
# segment_features['timepos']
# print most_similar_segment_index
# print segment_features['segpos'][most_similar_segment_index]
most_similar_timestamp = segment_features['timepos'][most_similar_segment_index]
# print most_similar_timestamp
# quit the [[[]]] (we got a nested array for some reason)
most_similar_timestamp = most_similar_timestamp[0][0]
# print most_similar_timestamp
# return tuple (start_pos, end_pos)
return (most_similar_timestamp[0],most_similar_timestamp[1])
def get_feature_timbre_ssd(wav_dir):
audiofile = wav_dir
samplerate, samplewidth, wavedata = audiofile_read(audiofile)
nsamples = wavedata.shape[0]
nchannels = wavedata.shape[1]
print "Successfully read audio file:", audiofile
print samplerate, "Hz,", samplewidth * 8, "bit,", nchannels, "channel(s),", nsamples, "samples"
features = rp_extract(wavedata, # the two-channel wave-data of the audio-file
samplerate=11025, # the samplerate of the audio-file
extract_ssd=True, # <== extract this feature!
transform_db=True, # apply psycho-accoustic transformation
transform_phon=True, # apply psycho-accoustic transformation
transform_sone=True, # apply psycho-accoustic transformation
fluctuation_strength_weighting=True, # apply psycho-accoustic transformation
skip_leadin_fadeout=1, # skip lead-in/fade-out. value = number of segments skipped
step_width=1) #
# plotssd(features['ssd'])
print len(features['ssd'])
return features
def extract_all_files(filelist,
path,
out_file=None,
feature_types=['rp', 'ssd', 'rh'],
label=False,
append=False,
no_extension_check=False,
force_resampling=None,
out_HDF5=False,
log_AudioTypes=True,
log_Errors=True,
verbose=True):
"""
finds all files of a certain type (e.g. .wav and/or .mp3) in a path and all sub-directories in it
extracts selected RP feature types
and saves them into separate CSV feature files (one per feature type)
# filelist: list of files for features to be extracted
# path: absolute path that will be added at beginning of filelist (can be '')
# out_file: output file name stub for feature files to write (if omitted, features will be returned from function)
# feature_types: RP feature types to extract. see rp_extract.py
# label: use subdirectory name as class label
# no_extension_check: does not check file format via extension. means that decoder is called on ALL files.
# force_resampling: force a target sampling rate (provided in Hz) when decoding (works with FFMPEG only!)
# out_HDF5: whether to store as HDF5 file format (otherwise CSV)
"""
ext = feature_types
n = 0 # counting the files being processed
n_extracted = 0 # counting the files that were actually analyzed
err = 0 # counting errors
n_files = len(filelist)
# initialize filelist_extracted and dict containing all accumulated feature arrays
filelist_extracted = []
feat_array = {}
audio_logwriter = None
error_logwriter = None
audio_logwriter_wrote_header = False
start_time = time.time()
if out_file: # only if out_file is specified
if log_AudioTypes:
log_filename = out_file + '.audiotypes.log'
audio_logfile = open(log_filename,
'w') # TODO allow append mode 'a'
audio_logwriter = unicsv.UnicodeCSVWriter(
audio_logfile) #, quoting=csv.QUOTE_ALL)
if log_Errors:
err_log_filename = out_file + '.errors.log'
error_logfile = open(err_log_filename,
'w') # TODO allow append mode 'a'
error_logwriter = unicsv.UnicodeCSVWriter(
error_logfile) #, quoting=csv.QUOTE_ALL)
if out_HDF5:
FeatureWriter = HDF5FeatureWriter()
else:
FeatureWriter = CSVFeatureWriter()
FeatureWriter.open(out_file, ext, append=append)
for fil in filelist: # iterate over all files
try:
if n > 0:
elaps_time = time.time() - start_time
remain_time = elaps_time * n_files / n - elaps_time # n is the number of files done here
else:
remain_time = None
n += 1
if path:
filename = path + os.sep + fil
else:
filename = fil
if verbose:
print '#', n, '/', n_files, '(ETA: ' + timestr(
remain_time) + "):", filename
# read audio file (wav or mp3)
samplerate, samplewidth, data, decoder = audiofile_read(
filename,
verbose=verbose,
include_decoder=True,
no_extension_check=no_extension_check,
force_resampling=force_resampling)
# audio file info
if verbose:
print samplerate, "Hz,", data.shape[
1], "channel(s),", data.shape[0], "samples"
# extract features
# Note: the True/False flags are determined by checking if a feature is listed in 'ext' (see settings above)
feat = rp.rp_extract(
data,
samplerate,
extract_rp=('rp' in ext), # extract Rhythm Patterns features
extract_ssd=(
'ssd' in ext), # extract Statistical Spectrum Descriptor
extract_tssd=(
'tssd' in ext
), # extract temporal Statistical Spectrum Descriptor
extract_rh=('rh' in ext), # extract Rhythm Histogram features
extract_trh=(
'trh'
in ext), # extract temporal Rhythm Histogram features
extract_mvd=(
'mvd' in ext
), # extract Modulation Frequency Variance Descriptor
spectral_masking=True,
transform_db=True,
transform_phon=True,
transform_sone=True,
fluctuation_strength_weighting=True,
skip_leadin_fadeout=1,
step_width=1,
verbose=verbose)
# TODO check if ext and feat.keys are consistent
# WHAT TO USE AS ID (based on filename): 3 choices:
id = fil # rel. filename as from find_files
# id = filename # full filename incl. full path
# id = filename[len(path)+1:] # relative filename only (extracted from path)
if out_file:
# WRITE each feature set to a CSV or HDF5 file
id2 = None
if label:
id2 = id.replace("\\", "/").split("/")[-2].strip()
if out_HDF5 and n_extracted == 0:
# for HDF5 we need to know the vector dimension
# thats why we cannot open the file earlier
FeatureWriter.open(
out_file, ext, feat, append=append
) # append not working for now but possibly in future
FeatureWriter.write_features(id, feat, id2)
else:
# IN MEMORY: add the extracted features for 1 file to the array dict accumulating all files
# TODO: only if we don't have out_file? maybe we want this as a general option
if feat_array == {}: # for first file, initialize empty array with dimension of the feature set
for e in feat.keys():
feat_array[e] = np.empty((0, feat[e].shape[0]))
# store features in array
for e in feat.keys():
feat_array[e] = np.append(
feat_array[e], feat[e].reshape(1, -1), axis=0
) # 1 for horizontal vector, -1 means take original dimension
filelist_extracted.append(id)
n_extracted += 1
# write list of analyzed audio files alongsize audio metadata (kHz, bit, etc.)
if audio_logwriter:
if not audio_logwriter_wrote_header: # write CSV header
log_info = [
"filename", "decoder", "samplerate (kHz)",
"samplewidth (bit)", "n channels", "n samples"
]
audio_logwriter.writerow(log_info)
audio_logwriter_wrote_header = True
log_info = [
filename, decoder, samplerate, samplewidth * 8,
data.shape[1], data.shape[0]
]
audio_logwriter.writerow(log_info)
gc.collect(
) # after every file we do garbage collection, otherwise our memory is used up quickly for some reason
except Exception as e:
print "ERROR analysing file: " + fil + ": " + str(e)
err += 1
if error_logwriter:
error_logwriter.writerow([fil, str(e)])
try:
if out_file: # close all output files
FeatureWriter.close()
if audio_logwriter:
audio_logfile.close()
if error_logwriter:
error_logfile.close()
except Exception as e:
print "ERROR closing the output or log files: " + str(e)
end_time = time.time()
if verbose:
print "FEATURE EXTRACTION FINISHED.", n, "file(s) processed,", n_extracted, "successful. Duration:", timestr(
end_time - start_time)
if err > 0:
print err, "file(s) had ERRORs during feature extraction.",
if log_Errors:
print "See", err_log_filename
else:
print
if out_file:
opt_ext = '.h5' if out_HDF5 else ''
print "Feature file(s):", out_file + "." + str(ext) + opt_ext
if out_file is None:
return filelist_extracted, feat_array
def extract_all_files_in_path(path,out_file,feature_types,audiofile_types=('.wav','.mp3')):
ext = feature_types
# get file list of all files in a path (filtered by audiofile_types)
filelist = find_files(path,audiofile_types,relative_path=True)
n = 0 # counting the files that were actually analyzed
err = 0 # counting errors
n_files = len(filelist)
start_abs = time.time()
files, writer = initialize_feature_files(out_file,ext)
for fil in filelist: # iterate over all files
try:
n += 1
filename = path + os.sep + fil
print '#',n,'/',n_files,':', filename
start = time.time()
# read audio file (wav or mp3)
samplerate, samplewidth, data = audiofile_read(filename)
end = time.time()
print end - start, "sec"
# audio file info
print samplerate, "Hz,", data.shape[1], "channels,", data.shape[0], "samples"
# extract features
# Note: the True/False flags are determined by checking if a feature is listed in 'ext' (see settings above)
start = time.time()
feat = rp.rp_extract(data,
samplerate,
extract_rp = ('rp' in ext), # extract Rhythm Patterns features
extract_ssd = ('ssd' in ext), # extract Statistical Spectrum Descriptor
extract_sh = ('sh' in ext), # extract Statistical Histograms
extract_tssd = ('tssd' in ext), # extract temporal Statistical Spectrum Descriptor
extract_rh = ('rh' in ext), # extract Rhythm Histogram features
extract_trh = ('trh' in ext), # extract temporal Rhythm Histogram features
extract_mvd = ('mvd' in ext), # extract Modulation Frequency Variance Descriptor
spectral_masking=True,
transform_db=True,
transform_phon=True,
transform_sone=True,
fluctuation_strength_weighting=True,
skip_leadin_fadeout=1,
step_width=1)
end = time.time()
print "Features extracted:", feat.keys(), end - start, "sec"
# WRITE each feature set to a CSV
# TODO check if ext and feat.keys are consistent
start = time.time()
# add filename before vector. 3 choices:
id = fil # filename only
# id = filename # full filename incl. full path
# id = filename[len(path)+1:] # relative filename only
write_feature_files(id,feat,writer)
end = time.time()
print "Data written." #, end-start
except:
print "ERROR analysing file: " + fil
err += 1
# close all output files
close_feature_files(files,ext)
end = time.time()
print "FEATURE EXTRACTION FINISHED.", n, "files,", end-start_abs, "sec"
if err > 0:
print err, "files had ERRORs during feature extraction."
if len(sys.argv) > 1:
if sys.argv[1] == '-test': # RUN DOCSTRING SELF TEST
print("Doing self test. If nothing is printed, it is ok.")
import doctest
doctest.run_docstring_examples(rp_extract,
globals()) #, verbose=True)
exit() # Note: no output means that everything went fine
else:
audiofile = sys.argv[1]
else:
audiofile = "music/BoxCat_Games_-_10_-_Epic_Song.mp3"
# Read audio file and extract features
try:
samplerate, samplewidth, wavedata = audiofile_read(audiofile)
np.set_printoptions(suppress=True)
bark_bands = 24 # choose the number of Bark bands (2..24)
mod_ampl_limit = 60 # number modulation frequencies on x-axis
feat = rp_extract(wavedata,
samplerate,
extract_rp=True,
extract_ssd=True,
extract_tssd=False,
extract_rh=True,
n_bark_bands=bark_bands,
spectral_masking=True,
transform_db=True,
# process file given on command line or default song (included)
if len(sys.argv) > 1:
if sys.argv[1] == '-test': # RUN DOCSTRING SELF TEST
print "Doing self test. If nothing is printed, it is ok."
import doctest
doctest.run_docstring_examples(rp_extract, globals()) #, verbose=True)
exit() # Note: no output means that everything went fine
else:
audiofile = sys.argv[1]
else:
audiofile = "music/BoxCat_Games_-_10_-_Epic_Song.mp3"
# Read audio file and extract features
try:
samplerate, samplewidth, wavedata = audiofile_read(audiofile)
np.set_printoptions(suppress=True)
bark_bands = 24 # choose the number of Bark bands (2..24)
mod_ampl_limit = 60 # number modulation frequencies on x-axis
feat = rp_extract(wavedata,
samplerate,
extract_rp=True,
extract_ssd=True,
extract_tssd=False,
extract_rh=True,
n_bark_bands=bark_bands,
spectral_masking=True,
transform_db=True,
def extract_all_files(filelist, path, out_file=None, feature_types=["rp", "ssd", "rh"], verbose=True):
"""
finds all files of a certain type (e.g. .wav and/or .mp3) in a path and all sub-directories in it
extracts selected RP feature types
and saves them into separate CSV feature files (one per feature type)
# filelist: list of files for features to be extracted
# path: absolute path that will be added at beginning of filelist (can be '')
# out_file: output file name stub for feature files to write (if omitted, features will be returned from function)
# feature_types: RP feature types to extract. see rp_extract.py
# audiofile_types: a string or tuple of suffixes to look for file extensions to consider (include the .)
"""
ext = feature_types
n = 0 # counting the files that were actually analyzed
err = 0 # counting errors
n_files = len(filelist)
# initialize filelist_extracted and dict containing all accumulated feature arrays
filelist_extracted = []
feat_array = {}
start_abs = time.time()
if out_file: # only if out_file is specified
files, writer = initialize_feature_files(out_file, ext)
for fil in filelist: # iterate over all files
try:
n += 1
if path:
filename = path + os.sep + fil
else:
filename = fil
# if verbose:
print "#", n, "/", n_files, ":", filename
# read audio file (wav or mp3)
samplerate, samplewidth, data = audiofile_read(filename)
# audio file info
if verbose:
print samplerate, "Hz,", data.shape[1], "channel(s),", data.shape[0], "samples"
# extract features
# Note: the True/False flags are determined by checking if a feature is listed in 'ext' (see settings above)
feat = rp.rp_extract(
data,
samplerate,
extract_rp=("rp" in ext), # extract Rhythm Patterns features
extract_ssd=("ssd" in ext), # extract Statistical Spectrum Descriptor
extract_sh=("sh" in ext), # extract Statistical Histograms
extract_tssd=("tssd" in ext), # extract temporal Statistical Spectrum Descriptor
extract_rh=("rh" in ext), # extract Rhythm Histogram features
extract_trh=("trh" in ext), # extract temporal Rhythm Histogram features
extract_mvd=("mvd" in ext), # extract Modulation Frequency Variance Descriptor
spectral_masking=True,
transform_db=True,
transform_phon=True,
transform_sone=True,
fluctuation_strength_weighting=True,
skip_leadin_fadeout=1,
step_width=1,
verbose=verbose,
)
# TODO check if ext and feat.keys are consistent
# WHAT TO USE AS ID (based on filename): 3 choices:
id = fil # rel. filename as from find_files
# id = filename # full filename incl. full path
# id = filename[len(path)+1:] # relative filename only (extracted from path)
if out_file:
# WRITE each feature set to a CSV
write_feature_files(id, feat, writer)
else:
# IN MEMORY: add the extracted features for 1 file to the array dict accumulating all files
# TODO: only if we dont have out_file? maybe we want this as a general option
if feat_array == {}: # for first file, initialize empty array with dimension of the feature set
for e in feat.keys():
feat_array[e] = np.empty((0, feat[e].shape[0]))
# store features in array
for e in feat.keys():
feat_array[e] = np.append(
feat_array[e], feat[e].reshape(1, -1), axis=0
) # 1 for horizontal vector, -1 means take original dimension
filelist_extracted.append(id)
except Exception as e:
print "ERROR analysing file: " + fil + ": " + str(e)
err += 1
if out_file: # close all output files
close_feature_files(files, ext)
end = time.time()
if verbose:
print "FEATURE EXTRACTION FINISHED. %d file(s), %.2f sec" % (n, end - start_abs)
if err > 0:
print err, "files had ERRORs during feature extraction."
if out_file:
print "Feature file(s):", out_file + ".*", ext
if out_file is None:
return filelist_extracted, feat_array
def extract_all_files(filelist,
path,
out_file=None,
feature_types=['rp', 'ssd', 'rh'],
verbose=True):
"""
finds all files of a certain type (e.g. .wav and/or .mp3) in a path and all sub-directories in it
extracts selected RP feature types
and saves them into separate CSV feature files (one per feature type)
# filelist: list of files for features to be extracted
# path: absolute path that will be added at beginning of filelist (can be '')
# out_file: output file name stub for feature files to write (if omitted, features will be returned from function)
# feature_types: RP feature types to extract. see rp_extract.py
# audiofile_types: a string or tuple of suffixes to look for file extensions to consider (include the .)
"""
ext = feature_types
n = 0 # counting the files that were actually analyzed
err = 0 # counting errors
n_files = len(filelist)
# initialize filelist_extracted and dict containing all accumulated feature arrays
filelist_extracted = []
feat_array = {}
start_abs = time.time()
if out_file: # only if out_file is specified
files, writer = initialize_feature_files(out_file, ext)
for fil in filelist: # iterate over all files
try:
n += 1
if path:
filename = path + os.sep + fil
else:
filename = fil
#if verbose:
print '#', n, '/', n_files, ':', filename
# read audio file (wav or mp3)
samplerate, samplewidth, data = audiofile_read(filename)
# audio file info
if verbose:
print samplerate, "Hz,", data.shape[
1], "channel(s),", data.shape[0], "samples"
# extract features
# Note: the True/False flags are determined by checking if a feature is listed in 'ext' (see settings above)
feat = rp.rp_extract(
data,
samplerate,
extract_rp=('rp' in ext), # extract Rhythm Patterns features
extract_ssd=(
'ssd' in ext), # extract Statistical Spectrum Descriptor
extract_sh=('sh' in ext), # extract Statistical Histograms
extract_tssd=(
'tssd' in ext
), # extract temporal Statistical Spectrum Descriptor
extract_rh=('rh' in ext), # extract Rhythm Histogram features
extract_trh=(
'trh'
in ext), # extract temporal Rhythm Histogram features
extract_mvd=(
'mvd' in ext
), # extract Modulation Frequency Variance Descriptor
spectral_masking=True,
transform_db=True,
transform_phon=True,
transform_sone=True,
fluctuation_strength_weighting=True,
skip_leadin_fadeout=1,
step_width=1,
verbose=verbose)
# TODO check if ext and feat.keys are consistent
# WHAT TO USE AS ID (based on filename): 3 choices:
id = fil # rel. filename as from find_files
# id = filename # full filename incl. full path
# id = filename[len(path)+1:] # relative filename only (extracted from path)
if out_file:
# WRITE each feature set to a CSV
write_feature_files(id, feat, writer)
else:
# IN MEMORY: add the extracted features for 1 file to the array dict accumulating all files
# TODO: only if we dont have out_file? maybe we want this as a general option
if feat_array == {}: # for first file, initialize empty array with dimension of the feature set
for e in feat.keys():
feat_array[e] = np.empty((0, feat[e].shape[0]))
# store features in array
for e in feat.keys():
feat_array[e] = np.append(
feat_array[e], feat[e].reshape(1, -1), axis=0
) # 1 for horizontal vector, -1 means take original dimension
filelist_extracted.append(id)
except Exception as e:
print "ERROR analysing file: " + fil + ": " + str(e)
err += 1
if out_file: # close all output files
close_feature_files(files, ext)
end = time.time()
if verbose:
print "FEATURE EXTRACTION FINISHED. %d file(s), %.2f sec" % (n, end -
start_abs)
if err > 0:
print err, "files had ERRORs during feature extraction."
if out_file: print "Feature file(s):", out_file + ".*", ext
if out_file is None:
return filelist_extracted, feat_array
def getSimilarSongs(audiofile, reference_db_filenames, reference_db_features, feature_type = 'rh'):
global query_features # GLOBAL VAR because it is reused in getSimilarSongSegments (see IMPORTANT note above)
# Read Wav File
samplerate, samplewidth, wavedata = audiofile_read(audiofile)
# Analyze New Audio
# here we set the feature type for the analysis, but we need to put it in a list [.]
fext = [feature_type]
query_features = rp_extract(wavedata,
samplerate,
extract_rp = ('rp' in fext), # extract Rhythm Patterns features
extract_ssd = ('ssd' in fext), # extract Statistical Spectrum Descriptor
extract_sh = ('sh' in fext), # extract Statistical Histograms
extract_tssd = ('tssd' in fext), # extract temporal Statistical Spectrum Descriptor
extract_rh = ('rh' in fext), # extract Rhythm Histogram features
extract_trh = ('trh' in fext), # extract temporal Rhythm Histogram features
extract_mvd = ('mvd' in fext), # extract Modulation Frequency Variance Descriptor
skip_leadin_fadeout=0,
step_width=1)
# from the dict returned by rp_extract, we take only the one feature_type we want
query_feature_vector = query_features[feature_type]
# reference_db_features should be a dict containing several feature types
# we take the feature type given as parameter
reference_feature_vectors = reference_db_features[feature_type]
# Search for similar songs in pre-analyzed song dataset (based on song's averaged features)
# initialize the similarity search object
sim_song_search = NearestNeighbors(n_neighbors = 6, metric='euclidean')
# TODO proper Scaling (Normalization) would need to add the live extracted song to the db_features
# Normalize the extracted features
# scaled_feature_space = StandardScaler().fit_transform(reference_feature_vectors)
# Fit the Nearest-Neighbor search object to the scaled features
# sim_song_search.fit(scaled_feature_space)
sim_song_search.fit(reference_feature_vectors)
# Get the most similar songs
(distances, similar_song_ids) = sim_song_search.kneighbors(query_feature_vector, return_distance=True)
# if the query is contained in the db feature set (usually if normalisation is used) we need to take it away from the result
#similar_songs = similar_songs[1:]
#print similar_song_ids
most_similar_songs = reference_db_filenames[feature_type][similar_song_ids]
# most_similar_songs is a 2D np.array, make it 1D
most_similar_songs = most_similar_songs[0]
# transform from np.array to simple list
most_similar_songs = most_similar_songs.tolist()
return(most_similar_songs)
