very bad quality of output inference

[TheAnkulai]

when i tried to inference I got this result with default settings
not_normal.wav
But with old rvc and same settings i got better result
good_result.wav
I dont know what's wrong, but that's sad
I made it on windows, rvc was used DirectML and inference started on cpu

[Hayri-maker]

Send me the datasetDen 8. aug. 2026 kl. 14.47 skrev TheAnkulai ***@***.***>:TheAnkulai created an issue (RVC-Project/Retrieval-based-Voice-Conversion-WebUI#2841) when i tried to inference I got this result with default settings not_normal.wav But with old rvc and same settings i got better result good_result.wav I dont know what's wrong, but that's sad —Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[TheAnkulai]

I don't know how, but it fixed by redownload and replace modules.py.
In first attempts on all methods I got this error (first lines was different):

[INFO]: device is not None, use cpu
  [INFO]    > call by:torchfcpe.tools.spawn_infer_cf_naive_mel_pe_from_pt
  [WARN] args.model.use_harmonic_emb is None; use default False
  [WARN]    > call by:torchfcpe.tools.spawn_cf_naive_mel_pe
Traceback (most recent call last):
  File "F:\Retrieval-based-Voice-Conversion-WebUI\infer\vc\modules.py", line 263, in vc_single
    audio_opt = self.pipeline.pipeline(
                ^^^^^^^^^^^^^^^^^^^^^^^
  File "F:\Retrieval-based-Voice-Conversion-WebUI\infer\vc\pipeline.py", line 328, in pipeline
    self.vc(
  File "F:\Retrieval-based-Voice-Conversion-WebUI\infer\vc\pipeline.py", line 222, in vc
    synthesized = run_cuda_graph(
                  ^^^^^^^^^^^^^^^
  File "F:\Retrieval-based-Voice-Conversion-WebUI\tools\cuda_graph.py", line 190, in run_cuda_graph
    return function(*inputs)
           ^^^^^^^^^^^^^^^^^
  File "F:\Retrieval-based-Voice-Conversion-WebUI\infer\vc\pipeline.py", line 225, in <lambda>
    lambda phone, lengths, coarse, continuous, speaker: net_g.infer(
                                                        ^^^^^^^^^^^^
  File "F:\Retrieval-based-Voice-Conversion-WebUI\infer\module\models.py", line 677, in infer
    o = self.dec(z * x_mask, nsff0, g=g, n_res=return_length2)
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "F:\Retrieval-based-Voice-Conversion-WebUI\.venv\Lib\site-packages\torch\nn\modules\module.py", line 1553, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "F:\Retrieval-based-Voice-Conversion-WebUI\.venv\Lib\site-packages\torch\nn\modules\module.py", line 1562, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "F:\Retrieval-based-Voice-Conversion-WebUI\infer\module\models.py", line 484, in forward
    har_source, noi_source, uv = self.m_source(f0, self.upp)
                                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "F:\Retrieval-based-Voice-Conversion-WebUI\.venv\Lib\site-packages\torch\nn\modules\module.py", line 1553, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "F:\Retrieval-based-Voice-Conversion-WebUI\.venv\Lib\site-packages\torch\nn\modules\module.py", line 1562, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "F:\Retrieval-based-Voice-Conversion-WebUI\infer\module\models.py", line 391, in forward
    sine_wavs, uv, _ = self.l_sin_gen(x, upp)
                       ^^^^^^^^^^^^^^^^^^^^^^
  File "F:\Retrieval-based-Voice-Conversion-WebUI\.venv\Lib\site-packages\torch\nn\modules\module.py", line 1553, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "F:\Retrieval-based-Voice-Conversion-WebUI\.venv\Lib\site-packages\torch\nn\modules\module.py", line 1562, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "F:\Retrieval-based-Voice-Conversion-WebUI\infer\module\models.py", line 335, in forward
    sine_waves = self._f02sine(f0, upp) * self.sine_amp
                 ^^^^^^^^^^^^^^^^^^^^^^
  File "F:\Retrieval-based-Voice-Conversion-WebUI\infer\module\models.py", line 315, in _f02sine
    rad += F.pad(rad_acc, (0, 0, 1, -1))
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "F:\Retrieval-based-Voice-Conversion-WebUI\.venv\Lib\site-packages\torch\nn\functional.py", line 4552, in pad
    return torch._C._nn.pad(input, pad, mode, value)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: torch-dml currently doesn't support a mix of positive and negative padding values, must be all positive or all negative.

But now after replace modules.py it starts work properly. So, at this time this issue can be considered irrelevant.
If you want to get .pth and .index that i used for this, there it is: https://dropmefiles.com/zm3vS (i can't upload zip in this message, idk why. And you able to download it only for 1 week). This model was trained not by me, so I don't know exactly dataset that was used for train. This is all I can give at this momentzz