verachen/Fun-ASR
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix #29

Browse Source 
Add support for low frame rate configuration in audio processing.
This commit is contained in:
彭震东
2025-12-23 14:09:45 +08:00
committed by GitHub
parent efe63c1229
commit 2b79dffe3e
1 changed files with 7 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
model.py
View File

@ -122,6 +122,7 @@ class FunASRNano(nn.Module):
param.requires_grad = False
audio_adaptor.eval()
self.audio_adaptor = audio_adaptor
self.use_low_frame_rate = audio_adaptor_conf.get("use_low_frame_rate", False)
self.length_normalized_loss = length_normalized_loss
self.feat_permute = audio_encoder_conf.get("feat_permute", True)
@ -387,9 +388,12 @@ class FunASRNano(nn.Module):
if self.feat_permute:
speech = speech.permute(0, 2, 1)
olens = 1 + (speech_lengths[0].item() - 3 + 2 * 1) // 2
olens = 1 + (olens - 3 + 2 * 1) // 2
fake_token_len_i = (olens - 1) // 2 + 1
if self.use_low_frame_rate:
olens = 1 + (speech_lengths[0].item() - 3 + 2 * 1) // 2
olens = 1 + (olens - 3 + 2 * 1) // 2
fake_token_len_i = (olens - 1) // 2 + 1
else:
fake_token_len_i = speech_lengths[0].item()
fake_token = [0] * fake_token_len_i
fbank_beg_i = len(source_ids)
source_ids += fake_token