diff --git a/autokara/autokara.py b/autokara/autokara.py
index ffb7edace7fad76076009591ed2a71cc06cc01b6..6077478b26f8f00229a6fdc88849705184440dd4 100644
--- a/autokara/autokara.py
+++ b/autokara/autokara.py
@@ -39,7 +39,8 @@ def main(opts=None):
'model': config['Segment']['model'],
'bdr': config['Segment'].getboolean('bdr'),
'cuda': config['Segment'].getboolean('cuda'),
- 'syl_delay': config['Segment'].getint('syl_delay')
+ 'syl_delay': config['Segment'].getint('syl_delay'),
+ 'uppercase_english': config['Segment'].getboolean('uppercase_as_english')
}
diff --git a/autokara/autosyl/LyricsAlignment/utils.py b/autokara/autosyl/LyricsAlignment/utils.py
index ec8aab853dbd5f016f5e5e9d55d020c7f93cdd7a..d6f996cc86a18f583585ce9655d4387943a4bbb7 100644
--- a/autokara/autosyl/LyricsAlignment/utils.py
+++ b/autokara/autosyl/LyricsAlignment/utils.py
@@ -7,6 +7,7 @@ import string
import warnings
import g2p_en
from g2p import make_g2p
+import re
@@ -110,11 +111,16 @@ def load_lyrics(lyrics_file):
def write_wav(path, audio, sr):
soundfile.write(path, audio.T, sr, "PCM_16")
-def gen_phone_gt(words, raw_lines, language="jp"):
+def gen_phone_gt(words, raw_lines, language="jp", uppercase_as_english=True):
- print(f"Translating lyrics to phonemes, language chosen : {language:s}")
+ print(f"Translating lyrics to phonemes, language chosen : {language:s}, uppercase as English : {str(uppercase_as_english):s}")
g2p = G2p_Wrapper(language=language)
+ if uppercase_as_english:
+ g2p_uppercase = G2p_Wrapper(language="en")
+ else:
+ g2p_uppercase = g2p
+ regex_uppercase = re.compile('[^a-z]')
# helper function
def getsubidx(x, y): # find y in x
@@ -125,7 +131,10 @@ def gen_phone_gt(words, raw_lines, language="jp"):
words_p = []
lyrics_p = []
for word in words:
- out = g2p(word)
+ if regex_uppercase.sub('', word) == '': # word is all uppercase or symbols
+ out = g2p_uppercase(word.lower())
+ else:
+ out = g2p(word.lower())
out = [phone if phone[-1] not in string.digits else phone[:-1] for phone in out]
words_p.append(out)
if len(lyrics_p) > 0:
@@ -146,7 +155,7 @@ def gen_phone_gt(words, raw_lines, language="jp"):
last_end = 0
for i in range(len(raw_lines)):
line = []
- line_phone = [g2p(word) for word in raw_lines[i].split()]
+ line_phone = [(g2p_uppercase(word.lower()) if regex_uppercase.sub('', word) == '' else g2p(word.lower())) for word in raw_lines[i].split()]
for l in line_phone:
line += l + [' ']
line = line[:-1]
diff --git a/autokara/autosyl/LyricsAlignment/wrapper.py b/autokara/autosyl/LyricsAlignment/wrapper.py
index f308ef9872955f4fdf4d45ce384c814fbee8664d..51cf853c1886c5def0ed49ca2d1341f67d05b3a1 100644
--- a/autokara/autosyl/LyricsAlignment/wrapper.py
+++ b/autokara/autosyl/LyricsAlignment/wrapper.py
@@ -10,10 +10,10 @@ from .model import train_audio_transforms, AcousticModel, BoundaryDetection
np.random.seed(7)
-def preprocess_from_file(audio_file, lyrics_file, word_file=None, language="jp"):
+def preprocess_from_file(audio_file, lyrics_file, word_file=None, language="jp", uppercase_as_english=True):
y, sr = preprocess_audio(audio_file)
- words, lyrics_p, idx_word_p, idx_line_p = preprocess_lyrics(lyrics_file, word_file, language=language)
+ words, lyrics_p, idx_word_p, idx_line_p = preprocess_lyrics(lyrics_file, word_file, language=language, uppercase_as_english=uppercase_as_english)
return y, words, lyrics_p, idx_word_p, idx_line_p
@@ -143,7 +143,7 @@ def preprocess_audio(audio_file, sr=22050):
return y, curr_sr
-def preprocess_lyrics(lyrics_lines, word_file=None, language="jp"):
+def preprocess_lyrics(lyrics_lines, word_file=None, language="jp", uppercase_as_english=True):
from string import ascii_lowercase
d = {ascii_lowercase[i]: i for i in range(26)}
d["'"] = 26
@@ -155,7 +155,7 @@ def preprocess_lyrics(lyrics_lines, word_file=None, language="jp"):
# raw_lines = f.read().splitlines()
raw_lines = lyrics_lines
- raw_lines = ["".join([c for c in line.lower() if c in d.keys()]).strip() for line in raw_lines]
+ raw_lines = ["".join([c for c in line if c.lower() in d.keys()]).strip() for line in raw_lines]
raw_lines = [" ".join(line.split()) for line in raw_lines if len(line) > 0]
# concat
full_lyrics = " ".join(raw_lines)
@@ -166,7 +166,7 @@ def preprocess_lyrics(lyrics_lines, word_file=None, language="jp"):
else:
words_lines = full_lyrics.split()
- lyrics_p, words_p, idx_word_p, idx_line_p = utils.gen_phone_gt(words_lines, raw_lines, language=(language if language else "jp"))
+ lyrics_p, words_p, idx_word_p, idx_line_p = utils.gen_phone_gt(words_lines, raw_lines, language=(language if language else "jp"), uppercase_as_english=uppercase_as_english)
return words_lines, lyrics_p, idx_word_p, idx_line_p
diff --git a/autokara/autosyl/segment.py b/autokara/autosyl/segment.py
index 46155073b8c10738a55de11981295909f38c91e9..0376de6c3cb83eca56f7d269d184692e1e4c0b4f 100644
--- a/autokara/autosyl/segment.py
+++ b/autokara/autosyl/segment.py
@@ -26,6 +26,7 @@ def segment(songfile, config, reference_syls=None,syls_per_line=10, last_syl_dur
checkpoint_folder = f"{str(Path(__file__).parent):s}/LyricsAlignment/checkpoints"
language = language
+ uppercase_as_english = config['uppercase_english']
lyrics_lines = [" ".join([syl[1] for syl in line]) for line in reference_syls]
#print(lyrics_lines)
@@ -38,7 +39,7 @@ def segment(songfile, config, reference_syls=None,syls_per_line=10, last_syl_dur
# lyrics_p: phoneme sequence of the target lyrics
# idx_word_p: indices of word start in lyrics_p
# idx_line_p: indices of line start in lyrics_p
- audio, words, lyrics_p, idx_word_p, idx_line_p = preprocess_from_file(audio_file, lyrics_lines, word_file, language)
+ audio, words, lyrics_p, idx_word_p, idx_line_p = preprocess_from_file(audio_file, lyrics_lines, word_file, language, uppercase_as_english=uppercase_as_english)
if verbose:
print(lyrics_p)
@@ -64,8 +65,20 @@ def segment(songfile, config, reference_syls=None,syls_per_line=10, last_syl_dur
magnitude = np.max(log_spec[:,:100], axis=1)
- magnitude_threshold = 0.75
- magnitude_start = 2
+ parsel = parselmouth.Sound(sig)
+
+ pitch = parsel.to_pitch()
+ pitch_values = pitch.selected_array['frequency']
+
+ pad_before = round(pitch.xs()[0]*100)
+ pad_after = len(magnitude) - len(pitch_values) - pad_before
+
+ pitch_values = np.pad(pitch_values, (pad_before, pad_after), 'constant', constant_values=(0,0))
+
+ mask = pitch_values
+
+ mask_threshold = 10
+ mask_start = 2
if reference_syls:
syls = []
@@ -73,9 +86,12 @@ def segment(songfile, config, reference_syls=None,syls_per_line=10, last_syl_dur
for line in reference_syls:
#print(onset_index, " : ", line)
l = [[onsets[onset_index + i], line[i][1]] for i in range(len(line)-1)]
- next_drop = words_onsets[onset_index + len(line) - 2] + magnitude_start
- while magnitude[next_drop] > magnitude_threshold:
+ next_drop = words_onsets[onset_index + len(line) - 2] + mask_start
+ after_high = False
+ while mask[next_drop] > mask_threshold or not after_high:
next_drop += 1
+ if mask[next_drop] > mask_threshold:
+ after_high = True
l.append([min(word_align[onset_index + (len(line) - 2)][1]/100, next_drop/100), ''])
syls.append(l)
onset_index += (len(line) - 1)
diff --git a/autokara/default.conf b/autokara/default.conf
index 82d914ce4a69e2792135de3cb620ee67f36e604e..4d48d0e3f361374a99fab8bfa5295c60ac3370a9 100644
--- a/autokara/default.conf
+++ b/autokara/default.conf
@@ -18,3 +18,5 @@ bdr = true
cuda = false
# Default delay applied to detected syls, in centiseconds
syl_delay = -4
+# Whether to treat uppercase words as English. If false, use song language everywhere
+uppercase_as_english = true
\ No newline at end of file
diff --git a/autokara/plot_syls.py b/autokara/plot_syls.py
index 9383639fc6cf3ab2a285eeaf9f8518ae8c53f9b5..b7c9ed46276c5601e2fec428da55cb411890a941 100644
--- a/autokara/plot_syls.py
+++ b/autokara/plot_syls.py
@@ -6,6 +6,7 @@ import matplotlib.pyplot as plt
import scipy.signal as sg
import parselmouth
import argparse
+from pathlib import Path
from .autosyl.assUtils import getSyls, timeToDate, dateToTime
from .autosyl.LyricsAlignment.wrapper import align, preprocess_from_file
@@ -43,8 +44,7 @@ def main(opts=None):
word_file = None # example: jamendolyrics/lyrics/*.words.txt"; Set to None if you don't have it
method = "MTL_BDR" # "Baseline", "MTL", "Baseline_BDR", "MTL_BDR"
cuda=True # set True if you have access to a GPU
- checkpoint_folder = "./autosyl/LyricsAlignment/checkpoints"
-
+ checkpoint_folder = f"{str(Path(__file__).parent):s}/autosyl/LyricsAlignment/checkpoints"
pred_file = "./MTL.csv" # saved alignment results, "(float) start_time, (float) end_time, (string) word"