actually pass language into dataset process script, fix coercing japanese into hiragana because espeak does not like kanji

scripts/process_dataset.py

@@ -185,7 +185,7 @@ for dataset_name in sorted(os.listdir(f'./{input_audio}/')):
 				try:
 					outpath, text, language, waveform, sample_rate = job
 
-					phones = valle_phonemize(text)
+					phones = valle_phonemize( text, language=language )
 					qnt = valle_quantize(waveform, sr=sample_rate, device=device)
 
 					if cfg.audio_backend == "dac":

vall_e/__main__.py

@@ -25,7 +25,7 @@ def main():
 	parser.add_argument("--input-prompt-length", type=float, default=3.0)
 
 	parser.add_argument("--top-p", type=float, default=1.0)
-	parser.add_argument("--top-k", type=int, default=16)
+	parser.add_argument("--top-k", type=int, default=0)
 	parser.add_argument("--repetition-penalty", type=float, default=1.0)
 	parser.add_argument("--repetition-penalty-decay", type=float, default=0.0)
 	parser.add_argument("--length-penalty", type=float, default=0.0)

vall_e/demo.py

@@ -56,7 +56,7 @@ def main():
 	parser.add_argument("--input-prompt-length", type=float, default=0.0)
 
 	parser.add_argument("--top-p", type=float, default=1.0)
-	parser.add_argument("--top-k", type=int, default=16)
+	parser.add_argument("--top-k", type=int, default=0)
 	parser.add_argument("--repetition-penalty", type=float, default=1.0)
 	parser.add_argument("--repetition-penalty-decay", type=float, default=0.0)
 	parser.add_argument("--length-penalty", type=float, default=0.0)
@@ -108,6 +108,8 @@ def main():
 
 	# pull from dataset samples
 	if args.sample_from_dataset:
+		cfg.dataset.cache = False
+
 		samples_dirs["dataset"] = args.demo_dir / "dataset"
 
 		print("Loading dataloader...")
@@ -157,13 +159,14 @@ def main():
 			text = open(dir / "prompt.txt").read()
 			language = open(dir / "language.txt").read() if (dir / "language.txt").exists() else "en"
 			prompt = dir / "prompt.wav"
+			reference = dir / "reference.wav"
 			out_path = dir / "out" / "ours.wav"
 
 			extra_sources = [ dir / "out" / f"{source}.wav" for source in sources ] if k == "librispeech" else []
 
 			samples.append((
 				text,
-			 	[ prompt, dir / "reference.wav", out_path ] + extra_sources
+			 	[ prompt, reference, out_path ] + extra_sources
 			))
 
 			if args.skip_existing and out_path.exists():

vall_e/emb/g2p.py

@@ -10,12 +10,22 @@ from phonemizer.backend import BACKENDS
 
 from tqdm import tqdm
 
+try:
+	import pykakasi
+except Exception as e:
+	pass
+
 @cache
 def _get_graphs(path):
 	with open(path, "r") as f:
 		graphs = f.read()
 	return graphs
 
+def romanize( runes, sep="" ):	
+	kks = pykakasi.kakasi()
+	result = kks.convert( runes )
+	return sep.join([ res['hira'] for res in result ])
+
 cached_backends = {}
 def _get_backend( language="en-us", backend="espeak" ):
 	key = f'{language}_{backend}'
@@ -37,6 +47,10 @@ def encode(text: str, language="en-us", backend="auto") -> list[str]:
 	if language == "en":
 		language = "en-us"
 
+	# Convert to kana because espeak does not like kanji...
+	if language[:2] == "ja" and backend == "auto":
+		text = romanize( text )
+
 	if not backend or backend == "auto":
 		backend = "espeak" # if language[:2] != "en" else "festival"
 
@@ -48,14 +62,7 @@ def encode(text: str, language="en-us", backend="auto") -> list[str]:
 	else:
 		tokens = phonemize( text, language=language, strip=True, preserve_punctuation=True, with_stress=True )
 	
+
 	tokens = list(tokens[0])
-	return tokens
-	"""
-	tokenized = " ".join( tokens )
 
-	merges = [ "\u02C8", "\u02CC", "\u02D0" ]
-	for merge in merges:
-		tokenized = tokenized.replace( f' {merge}', merge )
-
-	return tokenized.split(" ")
-	"""
+	return tokens