From 4e95ad3c4985dce64009ceeee70e30f9742418cd Mon Sep 17 00:00:00 2001
From: zhaomingwork <61895407+zhaomingwork@users.noreply.github.com>
Date: 星期四, 15 六月 2023 16:41:36 +0800
Subject: [PATCH] For python ws client long wav not recv asr result bug (#633)
---
funasr/models/e2e_asr_transducer.py | 8 +++++---
1 files changed, 5 insertions(+), 3 deletions(-)
diff --git a/funasr/models/e2e_asr_transducer.py b/funasr/models/e2e_asr_transducer.py
index 3120087..3f9f31c 100644
--- a/funasr/models/e2e_asr_transducer.py
+++ b/funasr/models/e2e_asr_transducer.py
@@ -108,7 +108,7 @@
self.use_auxiliary_lm_loss = auxiliary_lm_loss_weight > 0
if self.use_auxiliary_ctc:
- self.ctc_lin = torch.nn.Linear(encoder.output_size, vocab_size)
+ self.ctc_lin = torch.nn.Linear(encoder.output_size(), vocab_size)
self.ctc_dropout_rate = auxiliary_ctc_dropout_rate
if self.use_auxiliary_lm_loss:
@@ -162,7 +162,9 @@
# 1. Encoder
encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)
-
+ if hasattr(self.encoder, 'overlap_chunk_cls') and self.encoder.overlap_chunk_cls is not None:
+ encoder_out, encoder_out_lens = self.encoder.overlap_chunk_cls.remove_chunk(encoder_out, encoder_out_lens,
+ chunk_outs=None)
# 2. Transducer-related I/O preparation
decoder_in, target, t_len, u_len = get_transducer_task_io(
text,
@@ -577,7 +579,7 @@
self.use_auxiliary_lm_loss = auxiliary_lm_loss_weight > 0
if self.use_auxiliary_ctc:
- self.ctc_lin = torch.nn.Linear(encoder.output_size, vocab_size)
+ self.ctc_lin = torch.nn.Linear(encoder.output_size(), vocab_size)
self.ctc_dropout_rate = auxiliary_ctc_dropout_rate
if self.use_auxiliary_att:
--
Gitblit v1.9.1