From a2af08c32d96b136d3d91d28a6da0ba6ea52e00f Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期四, 15 六月 2023 17:10:12 +0800
Subject: [PATCH] Merge branch 'main' of github.com:alibaba-damo-academy/FunASR add
---
funasr/bin/asr_infer.py | 26 ++++++++++++++++++--------
1 files changed, 18 insertions(+), 8 deletions(-)
diff --git a/funasr/bin/asr_infer.py b/funasr/bin/asr_infer.py
index f9d6bf7..47ce0ee 100644
--- a/funasr/bin/asr_infer.py
+++ b/funasr/bin/asr_infer.py
@@ -493,9 +493,9 @@
else:
if pre_token_length[i] == 0:
yseq = torch.tensor(
- [self.asr_model.sos] + [self.asr_model.eos], device=yseq.device
+ [self.asr_model.sos] + [self.asr_model.eos], device=pre_acoustic_embeds.device
)
- score = torch.tensor(0.0, device=yseq.device)
+ score = torch.tensor(0.0, device=pre_acoustic_embeds.device)
else:
yseq = am_scores.argmax(dim=-1)
score = am_scores.max(dim=-1)[0]
@@ -1512,8 +1512,13 @@
if isinstance(speech, np.ndarray):
speech = torch.tensor(speech)
- feats = speech.unsqueeze(0).to(getattr(torch, self.dtype))
- feats_lengths = feats.new_full([1], dtype=torch.long, fill_value=feats.size(1))
+ if self.frontend is not None:
+ speech = torch.unsqueeze(speech, axis=0)
+ speech_lengths = speech.new_full([1], dtype=torch.long, fill_value=speech.size(1))
+ feats, feats_lengths = self.frontend(speech, speech_lengths)
+ else:
+ feats = speech.unsqueeze(0).to(getattr(torch, self.dtype))
+ feats_lengths = feats.new_full([1], dtype=torch.long, fill_value=feats.size(1))
if self.asr_model.normalize is not None:
feats, feats_lengths = self.asr_model.normalize(feats, feats_lengths)
@@ -1538,14 +1543,19 @@
if isinstance(speech, np.ndarray):
speech = torch.tensor(speech)
-
- feats = speech.unsqueeze(0).to(getattr(torch, self.dtype))
- feats_lengths = feats.new_full([1], dtype=torch.long, fill_value=feats.size(1))
+
+ if self.frontend is not None:
+ speech = torch.unsqueeze(speech, axis=0)
+ speech_lengths = speech.new_full([1], dtype=torch.long, fill_value=speech.size(1))
+ feats, feats_lengths = self.frontend(speech, speech_lengths)
+ else:
+ feats = speech.unsqueeze(0).to(getattr(torch, self.dtype))
+ feats_lengths = feats.new_full([1], dtype=torch.long, fill_value=feats.size(1))
feats = to_device(feats, device=self.device)
feats_lengths = to_device(feats_lengths, device=self.device)
- enc_out, _ = self.asr_model.encoder(feats, feats_lengths)
+ enc_out, _, _ = self.asr_model.encoder(feats, feats_lengths)
nbest_hyps = self.beam_search(enc_out[0])
--
Gitblit v1.9.1