dummy_input_length when exporting a vits model to onnx

[DarkoMilosevic86]

Hey everybody,
I was fixing the vits model export to onnx and thanks for accepting my pull request.
But, I have a help question about the dummy_input_length in the export_onnx function which is 100 by default

dummy_input_length = 100

When I try an exported onnx model using the onnx runtime, I'm getting only 2 sec of audio even if a text is bigger.
If I increase the dummy_input length, I AM getting a larger audio but with a noice at the end if a text is less.
Maybe I have a misstake in my python script for testing the onnx model, or I need to export a vits model to be dynammic, and I'm interesting how to do that.
MY test.py script is in the attachemt here.

Best regards,
Darko
test.py

[eginhard]

The value of dummy_input_length shouldn't have any effect on the exported model. I use this script for testing and there are no issues with longer audio:

# /// script
# requires-python = ">=3.12"
# dependencies = [
#     "coqui-tts[cpu,codec]",
#     "onnxruntime>=1.24.4",
#     "onnxscript>=0.6.2",
# ]
# ///
import numpy as np

from TTS.api import TTS
from TTS.tts.configs.vits_config import VitsConfig
from TTS.tts.models.vits import Vits
from TTS.utils.audio.numpy_transforms import save_wav

ONNX_FN = "coqui_vits.onnx"

tts = TTS("tts_models/en/ljspeech/vits")
tts.synthesizer.tts_model.export_onnx(verbose=False, output_path=ONNX_FN)
del tts

config_path = "~/.local/share/tts/tts_models--en--ljspeech--vits/config.json"

config = VitsConfig()
config.load_json(config_path)
vits = Vits(config)
vits.load_onnx(ONNX_FN)


text = "Hello world, this is the widely known voice of ljspeech dataset Hello world, this is the widely known voice of ljspeech dataset."

text_inputs = np.asarray(
    vits.tokenizer.text_to_ids(text, language=None),
    dtype=np.int64,
)[None, :]

audio = vits.inference_onnx(text_inputs)
save_wav(wav=audio[0], path="test_onnx_coqui.wav", sample_rate=config.audio.sample_rate)

[DarkoMilosevic86]

Thank you, it's works.