mirror of https://github.com/rhasspy/piper
You cannot select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
84 lines
2.3 KiB
Python
84 lines
2.3 KiB
Python
#!/usr/bin/env python3
|
|
import argparse
|
|
import json
|
|
import logging
|
|
import sys
|
|
import time
|
|
from pathlib import Path
|
|
|
|
import torch
|
|
|
|
from .vits.utils import audio_float_to_int16
|
|
from .vits.wavfile import write as write_wav
|
|
|
|
_LOGGER = logging.getLogger("piper_train.infer_generator")
|
|
|
|
|
|
def main():
|
|
"""Main entry point"""
|
|
logging.basicConfig(level=logging.DEBUG)
|
|
parser = argparse.ArgumentParser(prog="piper_train.infer_generator")
|
|
parser.add_argument("--model", required=True, help="Path to generator (.pt)")
|
|
parser.add_argument("--output-dir", required=True, help="Path to write WAV files")
|
|
parser.add_argument("--sample-rate", type=int, default=22050)
|
|
args = parser.parse_args()
|
|
|
|
args.output_dir = Path(args.output_dir)
|
|
args.output_dir.mkdir(parents=True, exist_ok=True)
|
|
|
|
model = torch.load(args.model)
|
|
|
|
# Inference only
|
|
model.eval()
|
|
|
|
for i, line in enumerate(sys.stdin):
|
|
line = line.strip()
|
|
if not line:
|
|
continue
|
|
|
|
utt = json.loads(line)
|
|
utt_id = str(i)
|
|
phoneme_ids = utt["phoneme_ids"]
|
|
speaker_id = utt.get("speaker_id")
|
|
|
|
text = torch.LongTensor(phoneme_ids).unsqueeze(0)
|
|
text_lengths = torch.LongTensor([len(phoneme_ids)])
|
|
sid = torch.LongTensor([speaker_id]) if speaker_id is not None else None
|
|
|
|
start_time = time.perf_counter()
|
|
audio = (
|
|
model(
|
|
text,
|
|
text_lengths,
|
|
sid,
|
|
# torch.FloatTensor([0.667]),
|
|
# torch.FloatTensor([1.0]),
|
|
# torch.FloatTensor([0.8]),
|
|
)[0]
|
|
.detach()
|
|
.numpy()
|
|
)
|
|
audio = audio_float_to_int16(audio)
|
|
end_time = time.perf_counter()
|
|
|
|
audio_duration_sec = audio.shape[-1] / args.sample_rate
|
|
infer_sec = end_time - start_time
|
|
real_time_factor = (
|
|
infer_sec / audio_duration_sec if audio_duration_sec > 0 else 0.0
|
|
)
|
|
|
|
_LOGGER.debug(
|
|
"Real-time factor for %s: %0.2f (infer=%0.2f sec, audio=%0.2f sec)",
|
|
i + 1,
|
|
real_time_factor,
|
|
infer_sec,
|
|
audio_duration_sec,
|
|
)
|
|
|
|
output_path = args.output_dir / f"{utt_id}.wav"
|
|
write_wav(str(output_path), args.sample_rate, audio)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|