|
|
import os.path |
|
|
import re |
|
|
|
|
|
from montreal_forced_aligner.command_line.mfa import mfa_cli |
|
|
from montreal_forced_aligner.config import TEMPORARY_DIRECTORY |
|
|
|
|
|
MODEL_VERSION = "3.2.0" |
|
|
|
|
|
root_dir = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) |
|
|
dictionary_dir = os.path.join(root_dir, "dictionary", "training") |
|
|
output_dir = os.path.join(root_dir, "g2p", "staging") |
|
|
temp_dir = TEMPORARY_DIRECTORY |
|
|
os.makedirs(output_dir, exist_ok=True) |
|
|
|
|
|
|
|
|
lang_codes = [ |
|
|
"czech", |
|
|
"russian", |
|
|
"french", |
|
|
"german", |
|
|
"portuguese_brazil", |
|
|
"portuguese_portugal", |
|
|
"spanish_spain", |
|
|
"spanish_latin_america", |
|
|
"swedish", |
|
|
"thai", |
|
|
"turkish", |
|
|
"english_us", |
|
|
"english_us_arpa", |
|
|
"english_uk", |
|
|
"english_nigeria", |
|
|
"korean", |
|
|
"hausa", |
|
|
"swahili", |
|
|
"vietnamese_hanoi", |
|
|
"vietnamese_hue", |
|
|
"vietnamese_ho_chi_minh_city", |
|
|
"ukrainian", |
|
|
"polish", |
|
|
"croatian", |
|
|
"bulgarian", |
|
|
"japanese", |
|
|
"japanese_katakana", |
|
|
|
|
|
"tamil", |
|
|
"hindi", |
|
|
"urdu", |
|
|
] |
|
|
|
|
|
lang_codes = [ |
|
|
"spanish_spain", |
|
|
"spanish_latin_america", |
|
|
"hindi", |
|
|
"urdu", |
|
|
|
|
|
|
|
|
] |
|
|
|
|
|
unicode_decomposition_languages = ["korean"] |
|
|
|
|
|
|
|
|
def get_error_rates(lang): |
|
|
train_temp_dir = os.path.join(temp_dir, f"{lang}_mfa") |
|
|
log_file = os.path.join(train_temp_dir, f"{lang}_mfa.log") |
|
|
print("Parsing:", log_file) |
|
|
if not os.path.exists(log_file): |
|
|
return 1, 1 |
|
|
wer_pattern = re.compile(r"WER:\s+([\d.]+)$") |
|
|
ler_pattern = re.compile(r"LER:\s+([\d.]+)$") |
|
|
wer, ler = None, None |
|
|
with open(log_file, "r", encoding="utf8") as f: |
|
|
for line in f: |
|
|
m = wer_pattern.search(line) |
|
|
if m: |
|
|
wer = m.group(0) |
|
|
m = ler_pattern.search(line) |
|
|
if m: |
|
|
ler = m.group(0) |
|
|
return wer, ler |
|
|
|
|
|
|
|
|
error_metrics = {} |
|
|
|
|
|
if __name__ == "__main__": |
|
|
for lang in lang_codes: |
|
|
print(lang) |
|
|
if lang.endswith("arpa"): |
|
|
dictionary_path = os.path.join(dictionary_dir, lang + ".dict") |
|
|
model_path = os.path.join(output_dir, lang + ".zip") |
|
|
else: |
|
|
dictionary_path = os.path.join(dictionary_dir, lang + "_mfa.dict") |
|
|
model_path = os.path.join(output_dir, lang + "_mfa.zip") |
|
|
print(model_path) |
|
|
if os.path.exists(model_path): |
|
|
error_metrics[lang] = get_error_rates(lang) |
|
|
continue |
|
|
unknown = [] |
|
|
if not os.path.exists(dictionary_path): |
|
|
continue |
|
|
|
|
|
command = [ |
|
|
"train_g2p", |
|
|
dictionary_path, |
|
|
model_path, |
|
|
"--clean", |
|
|
"-j", |
|
|
"10", |
|
|
"--use_mp", |
|
|
"--evaluate", |
|
|
"--num_pronunciations", |
|
|
"2", |
|
|
"--phonetisaurus", |
|
|
"--model_version", |
|
|
MODEL_VERSION, |
|
|
] |
|
|
if lang in unicode_decomposition_languages: |
|
|
command += ["--unicode_decomposition", "True"] |
|
|
mfa_cli(command, standalone_mode=False) |
|
|
|
|
|
error_metrics[lang] = get_error_rates(lang) |
|
|
|
|
|
for k, v in error_metrics.items(): |
|
|
print(f"{k}: {v[0]}% WER, {v[1]}% LER") |
|
|
|
