nodes/havvk-ComfyUI_AIIA
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

fix: remove int16 clipping and use float32 output in infer_v2.py

Browse Source
This commit is contained in:
Hawk Lee
2026-02-19 17:09:23 +08:00
parent f74b748eb8
commit 515d749177
1 changed files with 16 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
libs/index-tts/indextts/infer_v2.py
+16 -16
View File
@@ -656,21 +656,21 @@ class IndexTTS2:
s2mel_time += time.perf_counter() - m_start_time
m_start_time = time.perf_counter()
wav = self.bigvgan(vc_target.float()).squeeze().unsqueeze(0)
print(wav.shape)
bigvgan_time += time.perf_counter() - m_start_time
wav = wav.squeeze(1)
wav = self.bigvgan(vc_target.float()).squeeze().unsqueeze(0)
# print(wav.shape)
bigvgan_time += time.perf_counter() - m_start_time
wav = wav.squeeze(1)
wav = torch.clamp(32767 * wav, -32767.0, 32767.0)
if verbose:
print(f"wav shape: {wav.shape}", "min:", wav.min(), "max:", wav.max())
# wavs.append(wav[:, :-512])
wavs.append(wav.cpu()) # to cpu before saving
if stream_return:
yield wav.cpu()
if silence == None:
silence = self.interval_silence(wavs, sampling_rate=sampling_rate, interval_silence=interval_silence)
yield silence
# wav = torch.clamp(32767 * wav, -32767.0, 32767.0)
if verbose:
print(f"wav shape: {wav.shape}", "min:", wav.min(), "max:", wav.max())
# wavs.append(wav[:, :-512])
wavs.append(wav.cpu()) # to cpu before saving
if stream_return:
yield wav.cpu()
if silence == None:
silence = self.interval_silence(wavs, sampling_rate=sampling_rate, interval_silence=interval_silence)
yield silence
end_time = time.perf_counter()
self._set_gr_progress(0.9, "saving audio...")
@@ -694,7 +694,7 @@ class IndexTTS2:
print(">> remove old wav file:", output_path)
if os.path.dirname(output_path) != "":
os.makedirs(os.path.dirname(output_path), exist_ok=True)
torchaudio.save(output_path, wav.type(torch.int16), sampling_rate)
torchaudio.save(output_path, wav.float(), sampling_rate)
print(">> wav file saved to:", output_path)
if stream_return:
return None
@@ -703,7 +703,7 @@ class IndexTTS2:
if stream_return:
return None
# 返回以符合Gradio的格式要求
wav_data = wav.type(torch.int16)
wav_data = (wav * 32767).clamp(-32767, 32767).type(torch.int16)
wav_data = wav_data.numpy().T
yield (sampling_rate, wav_data)