From 493dda8f98662a546e908c08039898d0351782fd Mon Sep 17 00:00:00 2001
From: hnluo <haoneng.lhn@alibaba-inc.com>
Date: 星期四, 27 四月 2023 01:57:42 +0800
Subject: [PATCH] Update asr_inference_paraformer_streaming.py
---
funasr/bin/asr_inference_paraformer_streaming.py | 14 ++++++++------
1 files changed, 8 insertions(+), 6 deletions(-)
diff --git a/funasr/bin/asr_inference_paraformer_streaming.py b/funasr/bin/asr_inference_paraformer_streaming.py
index 939ffe9..c70baf0 100644
--- a/funasr/bin/asr_inference_paraformer_streaming.py
+++ b/funasr/bin/asr_inference_paraformer_streaming.py
@@ -202,10 +202,12 @@
assert check_argument_types()
results = []
cache_en = cache["encoder"]
- if speech.shape[1] < 16 * 60 and cache["is_final"]:
- cache["last_chunk"] = True
- feats = cache["feats"]
+ if speech.shape[1] < 16 * 60 and cache_en["is_final"]:
+ cache_en["tail_chunk"] = True
+ feats = cache_en["feats"]
feats_len = torch.tensor([feats.shape[1]])
+ results = self.infer(feats, feats_len, cache)
+ return results
else:
if self.frontend is not None:
feats, feats_len = self.frontend.forward(speech, speech_lengths, cache_en["is_final"])
@@ -232,7 +234,7 @@
feats_len = torch.tensor([feats_chunk2.shape[1]])
results_chunk2 = self.infer(feats_chunk2, feats_len, cache)
- return results_chunk1 + results_chunk2
+ return ["".join(results_chunk1 + results_chunk2)]
results = self.infer(feats, feats_len, cache)
@@ -466,7 +468,7 @@
cache_en = {"start_idx": 0, "cif_hidden": torch.zeros((batch_size, 1, 320)),
"cif_alphas": torch.zeros((batch_size, 1)), "chunk_size": chunk_size, "last_chunk": False,
- "feats": torch.zeros((batch_size, chunk_size[0] + chunk_size[2], 560))}
+ "feats": torch.zeros((batch_size, chunk_size[0] + chunk_size[2], 560)), "tail_chunk": False}
cache["encoder"] = cache_en
cache_de = {"decode_fsmn": None}
@@ -478,7 +480,7 @@
if len(cache) > 0:
cache_en = {"start_idx": 0, "cif_hidden": torch.zeros((batch_size, 1, 320)),
"cif_alphas": torch.zeros((batch_size, 1)), "chunk_size": chunk_size, "last_chunk": False,
- "feats": torch.zeros((batch_size, chunk_size[0] + chunk_size[2], 560))}
+ "feats": torch.zeros((batch_size, chunk_size[0] + chunk_size[2], 560)), "tail_chunk": False}
cache["encoder"] = cache_en
cache_de = {"decode_fsmn": None}
--
Gitblit v1.9.1