mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-08-13 20:12:24 +00:00
Code Issues Packages Projects Releases Wiki Activity

Minor fixes to the onnx inference script for ljspeech matcha-tts.

Browse Source
This commit is contained in:
Fangjun Kuang 2024-12-18 10:33:49 +08:00
parent 3e4da5f781
commit 2cf5e9a99c
2 changed files with 15 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
egs/ljspeech/TTS/matcha/export_onnx.py
View File

@ -163,7 +163,7 @@ def main():
(x, x_lengths, temperature, length_scale), (x, x_lengths, temperature, length_scale),
filename, filename,
opset_version=opset_version, opset_version=opset_version,
input_names=["x", "x_length", "temperature", "length_scale"], input_names=["x", "x_length", "noise_scale", "length_scale"],
output_names=["mel"], output_names=["mel"],
dynamic_axes={ dynamic_axes={
"x": {0: "N", 1: "L"}, "x": {0: "N", 1: "L"},

19
egs/ljspeech/TTS/matcha/onnx_pretrained.py
View File

@ -89,6 +89,7 @@ class OnnxHifiGANModel:
self.model.get_inputs()[0].name: x.numpy(), self.model.get_inputs()[0].name: x.numpy(),
}, },
)[0] )[0]
# audio: (batch_size, num_samples)
return torch.from_numpy(audio) return torch.from_numpy(audio)
@ -97,19 +98,24 @@ class OnnxModel:
def __init__( def __init__(
self, self,
filename: str, filename: str,
tokens: str,
): ):
session_opts = ort.SessionOptions() session_opts = ort.SessionOptions()
session_opts.inter_op_num_threads = 1 session_opts.inter_op_num_threads = 1
session_opts.intra_op_num_threads = 2 session_opts.intra_op_num_threads = 2
self.session_opts = session_opts self.session_opts = session_opts
self.tokenizer = Tokenizer("./data/tokens.txt") self.tokenizer = Tokenizer(tokens)
self.model = ort.InferenceSession( self.model = ort.InferenceSession(
filename, filename,
sess_options=self.session_opts, sess_options=self.session_opts,
providers=["CPUExecutionProvider"], providers=["CPUExecutionProvider"],
) )
logging.info(f"{self.model.get_modelmeta().custom_metadata_map}")
metadata = self.model.get_modelmeta().custom_metadata_map
self.sample_rate = int(metadata["sample_rate"])
for i in self.model.get_inputs(): for i in self.model.get_inputs():
print(i) print(i)
@ -138,6 +144,7 @@ class OnnxModel:
self.model.get_inputs()[3].name: length_scale.numpy(), self.model.get_inputs()[3].name: length_scale.numpy(),
}, },
)[0] )[0]
# mel: (batch_size, feat_dim, num_frames)
return torch.from_numpy(mel) return torch.from_numpy(mel)
@ -147,7 +154,7 @@ def main():
params = get_parser().parse_args() params = get_parser().parse_args()
logging.info(vars(params)) logging.info(vars(params))
model = OnnxModel(params.acoustic_model) model = OnnxModel(params.acoustic_model, tokens)
vocoder = OnnxHifiGANModel(params.vocoder) vocoder = OnnxHifiGANModel(params.vocoder)
text = params.input_text text = params.input_text
x = model.tokenizer.texts_to_token_ids([text], add_sos=True, add_eos=True) x = model.tokenizer.texts_to_token_ids([text], add_sos=True, add_eos=True)
@ -164,15 +171,17 @@ def main():
print("audio", audio.shape) # (1, 1, num_samples) print("audio", audio.shape) # (1, 1, num_samples)
audio = audio.squeeze() audio = audio.squeeze()
sample_rate = model.sample_rate
t = (end_t - start_t).total_seconds() t = (end_t - start_t).total_seconds()
t2 = (end_t2 - start_t2).total_seconds() t2 = (end_t2 - start_t2).total_seconds()
rtf_am = t * 22050 / audio.shape[-1] rtf_am = t * sample_rate / audio.shape[-1]
rtf_vocoder = t2 * 22050 / audio.shape[-1] rtf_vocoder = t2 * sample_rate / audio.shape[-1]
print("RTF for acoustic model ", rtf_am) print("RTF for acoustic model ", rtf_am)
print("RTF for vocoder", rtf_vocoder) print("RTF for vocoder", rtf_vocoder)
# skip denoiser # skip denoiser
sf.write(params.output_wav, audio, 22050, "PCM_16") sf.write(params.output_wav, audio, sample_rate, "PCM_16")
logging.info(f"Saved to {params.output_wav}") logging.info(f"Saved to {params.output_wav}")