def train_2_baidu():
out_file2 = codecs.open(_data_path + "/wav/S0002.txt",
'w',
encoding="utf-8")
for i, line in enumerate(
open(_data_path +
"/resource_aishell/aishell_transcript_v0.8.txt").readlines()):
# cfg = {
# 'spd': random.randint(4, 6), # 语速，取值0-9，默认为5中语速
# 'pit': random.randint(4, 7), # 音调，取值0-9，默认为5中语调
# 'vol': random.randint(4, 7), # 音量，取值0-15，默认为5中音量
# 'per': random.randint(0, 3) # 发音人选择, 0为女声，1为男声，3为情感合成-度逍遥，4为情感合成-度丫丫(不好)，默认为普通女
# }
file1, d = line.strip().split(" ", 1)
if file1 >= "BAC009S0003W0121":
continue
cfg = {'spd': 5, 'pit': 5, 'vol': 5, 'per': 0}
# d = json.loads(line.strip())
path = os.path.join(
_data_path, "wav", file1[6:11],
"%s_%s_%s_%s_%s.%s" % (file1, cfg.get("spd"), cfg.get("pit"),
cfg.get("vol"), cfg.get("per"), "mp3"))
gen_wav(d, cfg, path)
AudioSegment.from_mp3(path).export(path[:-3] + "wav", format="wav")
ps = generate_zi_label(d)
lin = "{\"key\":\"" + path[:-3] + "wav" + "\", \"duration\": " + str(
get_duration_wave(path[:-3] + "wav")) + ", \"text\":\"" + " ".join(
ps).decode("utf-8") + "\"}"
out_file2.write(lin + "\n")
if (i + 1) % 100 == 0:
out_file2.flush()
out_file2.close()
def xiaoshuo_2_word():
d = set()
for i, line in enumerate(open("resources/unicodemap_zi.csv").readlines()):
d.add(line.rsplit(",", 1)[0])
DIR = "/export/aiplatform/8k/"
out_file = open(DIR + 'resulttxtnew26.json', 'w')
for i in glob.glob(DIR + "resulttxtnew26/*/*.wav"):
txt = "".join([
line.strip()
for line in open(i.replace("8k/", "")[:-3] + "txt").readlines()
])
txt = strQ2B(txt.strip().decode("utf8")).encode("utf8")
ps = generate_zi_label(deletePunc(txt))
if len(ps) == 0:
continue
flag = False
for p in ps:
if p not in d or p.isdigit():
print("not in d is %s %s. %s" % (p, [p], "".join(ps)))
flag = True
break
if flag:
continue
duration = get_duration_wave(i)
if duration > 16:
continue
line = "{\"key\":\"" + i + "\", \"duration\": " + str(
duration) + ", \"text\":\"" + " ".join(ps) + "\"}"
out_file.write(line + "\n")
out_file.close()
def client_2_word():
DIR = "/export/aiplatform/client_files4/"
def compare(x, y):
stat_x = os.stat(DIR + "/" + x)
stat_y = os.stat(DIR + "/" + y)
if stat_x.st_ctime < stat_y.st_ctime:
return -1
elif stat_x.st_ctime > stat_y.st_ctime:
return 1
else:
return 0
# iterms = os.listdir(DIR)
# iterms.sort(compare)
# for iterm in iterms:
# print(iterm)
wavs = open(DIR + 'wav.txt').readlines()
labels = open(DIR + 'label.txt').readlines()
out_file = open(DIR + 'client4.json', 'w')
for i, (path, txt) in enumerate(zip(wavs, labels)):
ps = generate_zi_label(
txt.replace(",", "").replace("。", "").replace("，", "").strip())
audio_path = DIR + path.strip()
duration = get_duration_wave(audio_path)
line = "{\"key\":\"" + audio_path + "\", \"duration\": " + str(
duration) + ", \"text\":\"" + " ".join(ps) + "\"}"
out_file.write(line + "\n")
out_file.close()
def gen_label():
f = open(_data_path + "001_baidu.txt")
# wfobj = codecs.open(_data_path + "001_baidu_2.txt", 'w', encoding="utf-8")
wfobj = open(_data_path + "001_baidu_2.txt", 'w')
for i, line in enumerate(f.readlines()):
if not line.strip():
continue
newi = generate_zi_label(deletePunc(line.strip()))
wfobj.write(" ".join(newi) + "\n")
if i % 10 == 0:
wfobj.flush()
wfobj.close()
def search_2_word():
tran = open("/export/aiplatform/search/transcript").readlines()
out_file = open('resources/search.json', 'w')
for t in tran:
path, d, txt = t.split(" ", 2)
ps = generate_zi_label(txt.strip())
audio_path = "/export/aiplatform/search/" + "wav/" + path
duration = get_duration_wave(audio_path)
line = "{\"key\":\"" + audio_path + "\", \"duration\": " + str(
duration) + ", \"text\":\"" + " ".join(ps) + "\"}"
out_file.write(line + "\n")
out_file.close()
def aia_2_word(DIR):
scp = [i for i in glob.glob(DIR + "/*/*.scp") if "noise" not in i]
dir_name = DIR.rsplit("/", 1)[1]
out_file = codecs.open(_data_path + 'fanlu/' + dir_name + '.json',
'w',
encoding="utf-8")
d = set()
e = set()
for i, line in enumerate(open("resources/unicodemap_zi.csv").readlines()):
d.add(line.rsplit(",", 1)[0])
out_file2 = codecs.open(_data_path + 'fanlu/' + dir_name + '.miss.json',
'w',
encoding="utf-8")
for j in scp:
for m, line in enumerate(open(j).readlines()):
# print(line)
file_name, txt = line.strip().split("\t", 1)
path = "/export/fanlu/" + '16k/' + dir_name + "/" + j.rsplit(
"/", 1)[1].replace(".scp", "") + "/" + file_name + ".wav"
if not os.path.exists(path):
print("%s not exist" % path)
continue
duration = get_duration_wave(path)
if duration > 16:
print("%s longer than 16s" % path)
continue
txt = strQ2B(txt.strip().decode("utf8")).encode("utf8")
ps = generate_zi_label(deletePunc(txt))
if len(ps) == 0:
continue
line = "{\"key\":\"" + path.replace(
"fanlu", "aiplatform") + "\", \"duration\": " + str(
duration) + ", \"text\":\"" + " ".join(
[p.decode("utf-8") for p in ps]) + "\"}"
flag = False
for p in ps:
if p not in d:
e.add(p)
print("not in d is %s %s. %s" % (p, [p], "".join(ps)))
flag = True
break
if flag:
out_file2.write(line + "\n")
else:
out_file.write(line + "\n")
out_file.close()
out_file2.close()
out_file1 = open(_data_path + 'fanlu/' + dir_name + '.miss', 'w')
for i in e:
out_file1.write(i + "\n")
out_file1.close()
def check_biaozhu():
f = _data_path + "bdp1.txt"
import json
count = 0
all = 0
amount = 0
import codecs
wfobj = codecs.open(_data_path + "bdp2.txt", 'w', encoding="utf-8")
# f2 = open(_data_path + "bdp2.txt", "w")
for i in open(f).readlines():
d = json.loads(i.strip())
manual = d.get("manual", "").encode("utf-8").replace("，", "").replace(
"。", "").replace(",", "").replace(".", "")
machine = d.get("machine", "").encode("utf-8").replace(
"，", "").replace("。", "").replace(",", "").replace(".", "")
if "A" not in manual and "B" not in manual and "C" not in manual and "D" not in manual and "E" not in manual:
manuals = generate_zi_label(manual)
machines = generate_zi_label(machine)
l_distance = levenshtein_distance(manuals, machines)
count += l_distance
all += len(manuals)
amount += 1
wav_file = "/export/aiplatform/data_label/task0/" + d.get(
"name", "")
duration = get_duration_wave(wav_file)
if duration > 16:
continue
c = {
"key": wav_file,
"duration": str(duration),
"text": " ".join([m.decode("utf-8") for m in manuals])
}
# line = "{\"key\":\"" + wav_file + "\", \"duration\": " + str(1) + ", \"text\":\"" + " ".join([m.decode("utf-8") for m in manuals]) + "\"}"
wfobj.write(json.dumps(c, ensure_ascii=False) + "\n")
wfobj.close()
print("amount: %d, error: %d, all: %d, cer: %.4f" %
(amount, count, all, count / float(all)))
def ai_thchs30_2_word():
ori_wavs = glob.glob(_data_path + "/thchs30/data_thchs30/8k/data/*.wav")
out_file = open(_data_path + "/thchs30/data_thchs30/8k/thchs30_data.json",
'w')
for w in ori_wavs:
path, name = w.rsplit("/", 1)
rs = open(w.replace("8k", "") + ".trn").readlines()[0].strip()
ps = generate_zi_label(rs)
duration = get_duration_wave(w)
line = "{\"key\":\"" + w + "\", \"duration\": " + str(
duration) + ", \"text\":\"" + " ".join(ps) + "\"}"
out_file.write(line + "\n")
# for w2 in glob.glob(path.replace("data","data_aug")+"/" + name.split(".")[0] + "*.wav"):
# audio = wave.open(w2)
# duration = float(audio.getnframes()) / audio.getframerate()
# audio.close()
# line = "{\"key\":\"" + w2 + "\", \"duration\": " + str(duration) + ", \"text\":\"" + " ".join(ps) + "\"}"
# out_file.write(line + "\n")
# print(ps)
out_file.close()
def ai_2_word():
lines = open(
_data_path +
"data_aishell/transcript/aishell_transcript_v0.8.txt").readlines()
out_file = open(_data_path + "data_aishell/wav8000/aishell_train_8k.json",
'w')
out_file1 = open(
_data_path + "data_aishell/wav8000/aishell_validation_8k.json", 'w')
out_file2 = open(_data_path + "data_aishell/wav8000/aishell_test_8k.json",
'w')
for line in lines:
rs = line.strip().split(" ")
ps = generate_zi_label("".join(rs[1:]))
if rs[0][6:11] <= "S0723":
wav = _data_path + "data_aishell/wav8000/train/" + rs[0][
6:11] + "/" + rs[0] + ".wav"
# dir = _data_path + "data_aishell/wav/train_aug/" + rs[0][6:11] + "/" + rs[0]
# for w in glob.glob(dir + "*.wav"):
duration = get_duration_wave(wav)
line = "{\"key\":\"" + wav + "\", \"duration\": " + str(
duration) + ", \"text\":\"" + " ".join(ps) + "\"}"
out_file.write(line + "\n")
elif rs[0][6:11] <= "S0763":
wav = _data_path + "data_aishell/wav8000/dev/" + rs[0][
6:11] + "/" + rs[0] + ".wav"
duration = get_duration_wave(wav)
line = "{\"key\":\"" + wav + "\", \"duration\": " + str(
duration) + ", \"text\":\"" + " ".join(ps) + "\"}"
out_file1.write(line + "\n")
else:
wav = _data_path + "data_aishell/wav/test/" + rs[0][
6:11] + "/" + rs[0] + ".wav"
duration = get_duration_wave(wav)
line = "{\"key\":\"" + wav + "\", \"duration\": " + str(
duration) + ", \"text\":\"" + " ".join(ps) + "\"}"
out_file2.write(line + "\n")
out_file.close()
out_file1.close()
out_file2.close()
def deal_wave():
d = set()
for i, line in enumerate(open("resources/unicodemap_zi.csv").readlines()):
d.add(line.rsplit(",", 1)[0])
out_file = open('/export/fanlu/WAVE.json', 'w')
script = glob.glob("/export/fanlu/SCRIPT/*.TXT")
for s in script:
name = s.rsplit("/", 1)[0][1:-5]
wav_path = "/export/file_server/WAVE/SPEAKER" + name + "/"
lines = open(s).readlines()
for i in range(0, len(lines), 2):
wav_name = lines[i].split("\t")[0]
txt = filter(is_chinese, lines[i + 1].strip().decode("utf-8"))
ps = generate_zi_label(txt)
if len(ps) == 0:
continue
flag = False
for p in ps:
if p not in d or p.isdigit():
print("not in d is %s %s. %s" % (p, [p], "".join(ps)))
flag = True
break
if flag:
continue
wav = wav_path + "SESSION0/" + wav_name + ".WAV"
if not os.path.exists(wav):
print("%s not exist" % wav)
continue
duration = get_duration_wave(wav)
if duration > 16:
print("%s longer than 16s" % wav)
continue
line = "{\"key\":\"" + wav.replace(
"file_server", "aiplatform") + "\", \"duration\": " + str(
duration) + ", \"text\":\"" + " ".join(ps) + "\"}"
out_file.write(line + "\n")
'spd': random.randint(4, 6), # 语速，取值0-9，默认为5中语速
'pit': random.randint(4, 7), # 音调，取值0-9，默认为5中语调
'vol': random.randint(4, 7), # 音量，取值0-15，默认为5中音量
'per': random.randint(
0, 3) # 发音人选择, 0为女声，1为男声，3为情感合成-度逍遥，4为情感合成-度丫丫(不好)，默认为普通女
}
file1, d = line.strip().split(" ", 1)
if file1 < 'BAC009S0208W0465':
continue
path = os.path.join(
_data_path, "data_aishell/baidu/", file1[6:11],
"%s_%s_%s_%s_%s.%s" % (file1, cfg.get("spd"), cfg.get("pit"),
cfg.get("vol"), cfg.get("per"), "mp3"))
gen_wav(d, cfg, path)
AudioSegment.from_mp3(path).export(path[:-3] + "wav", format="wav")
ps = generate_zi_label(d)
lin = "{\"key\":\"" + path + "\", \"duration\": " + str(
get_duration_wave(path[:-3] + "wav")) + ", \"text\":\"" + " ".join(
ps).decode("utf-8") + "\"}"
out_file2.write(lin + "\n")
if (i + 1) % 100 == 0:
print(i)
out_file2.flush()
out_file2.close()
# 从URL获取文件识别
# aipSpeech.asr('', 'pcm', 16000, {
# 'url': 'http://121.40.195.233/res/16k_test.pcm',
# 'callback': 'http://xxx.com/receive',
# })
def prepare_minibatch_fbank(self,
audio_paths,
texts,
overwrite=False,
is_bi_graphemes=False,
seq_length=-1,
save_feature_as_csvfile=False,
language="en",
zh_type="zi",
noise_percent=0.4):
""" Featurize a minibatch of audio, zero pad them and return a dictionary
Params:
audio_paths (list(str)): List of paths to audio files
texts (list(str)): List of texts corresponding to the audio files
Returns:
dict: See below for contents
"""
assert len(audio_paths) == len(texts), \
"Inputs and outputs to the network must be of the same number"
# Features is a list of (timesteps, feature_dim(161)) arrays (channel(3), feature_dim(41), timesteps)
# Calculate the features for each audio clip, as the log of the
# Fourier Transform of the audio
features = [
self.featurize_fbank(
a,
overwrite=overwrite,
save_feature_as_csvfile=save_feature_as_csvfile,
noise_percent=noise_percent,
seq_length=seq_length) for a in audio_paths
]
input_lengths = [f.shape[1] for f in features]
channel, timesteps, feature_dim = features[0].shape
mb_size = len(features)
# Pad all the inputs so that they are all the same length
if seq_length == -1:
x = np.zeros((mb_size, channel, self.max_seq_length, feature_dim))
else:
x = np.zeros((mb_size, channel, seq_length, feature_dim))
y = np.zeros((mb_size, self.max_label_length))
labelUtil = LabelUtil()
label_lengths = []
for i in range(mb_size):
feat = features[i]
feat = self.normalize_fbank(feat) # Center using means and std
x[i, :, :feat.
shape[1], :] = feat # padding with 0 padding with noise？
if language == "en" and is_bi_graphemes:
label = generate_bi_graphemes_label(texts[i])
label = labelUtil.convert_bi_graphemes_to_num(label)
y[i, :len(label)] = label
elif language == "en" and not is_bi_graphemes:
label = labelUtil.convert_word_to_num(texts[i])
y[i, :len(texts[i])] = label
elif language == "zh" and zh_type == "phone":
label = generate_phone_label(texts[i])
label = labelUtil.convert_bi_graphemes_to_num(label)
y[i, :len(label)] = label
elif language == "zh" and zh_type == "py":
label = generate_py_label(texts[i])
label = labelUtil.convert_bi_graphemes_to_num(label)
y[i, :len(label)] = label
elif language == "zh" and zh_type == "zi":
label = generate_zi_label(texts[i])
label = labelUtil.convert_bi_graphemes_to_num(label)
y[i, :len(label)] = label
label_lengths.append(len(label))
return {
'x': x, # (0-padded features of shape(mb_size,timesteps,feat_dim)
'y': y, # list(int) Flattened labels (integer sequences)
'texts': texts, # list(str) Original texts
'input_lengths': input_lengths, # list(int) Length of each input
'label_lengths': label_lengths, # list(int) Length of each label
}
