Added support for multiple speakers in mrs

datasets/mrs.py

@@ -7,7 +7,7 @@ import sqlite3
 import sys
 
 
-def build_from_path(in_dir, out_dir, num_workers=1, tqdm=lambda x: x):
+def build_from_path(in_dir, out_dir, username, num_workers=1, tqdm=lambda x: x):
   '''Preprocesses the recordings from mimic-recording-studio (based on sqlite db) into a given output directory.
 
     Args:
@@ -32,11 +32,19 @@ def build_from_path(in_dir, out_dir, num_workers=1, tqdm=lambda x: x):
   conn = sqlite3.connect(dbfile)
   c = conn.cursor()
 
-  # Get user id from sqlite to find recordings in directory structure
-  # TODO: Currently just works with one user
-  for row in c.execute('SELECT uuid FROM usermodel LIMIT 1;'):
+  uid = ''
+  sql_get_guid = "SELECT uuid FROM usermodel LIMIT 1;"
+  if username:
+    print("Query user guid for " + username + " in sqlite db")
+    sql_get_guid = "SELECT uuid FROM usermodel WHERE UPPER(user_name) = '" + username.upper() + "' LIMIT 1;"
+
+  for row in c.execute(sql_get_guid):
     uid = row[0]
   
+  if uid == '':
+    print("No userid could be found in sqlite db.")
+    sys.exit()
+  
   print("Found speaker user guid in sqlite: " + uid)
 
   wav_dir = os.path.join(in_dir,"backend","audio_files",uid)

preprocess.py

@@ -28,9 +28,10 @@ def preprocess_ljspeech(args):
 def preprocess_mrs(args):
   in_dir = args.mrs_dir
   out_dir = os.path.join(args.base_dir, args.output)
+  username = args.mrs_username
   os.makedirs(out_dir, exist_ok=True)
   metadata = mrs.build_from_path(
-      in_dir, out_dir, args.num_workers, tqdm=tqdm)
+      in_dir, out_dir, username, args.num_workers, tqdm=tqdm)
   write_metadata(metadata, out_dir)
 
 def preprocess_amy(args):
@@ -87,6 +88,7 @@ def main():
   parser = argparse.ArgumentParser()
   parser.add_argument('--base_dir', default=os.path.expanduser('~/tacotron'))
   parser.add_argument('--mrs_dir', required=False)
+  parser.add_argument('--mrs_username', required=False)
   parser.add_argument('--output', default='training')
   parser.add_argument(
       '--dataset', required=True, choices=['amy', 'blizzard', 'ljspeech', 'kusal', 'mailabs','mrs']