From 3e9319263835bd018abe2dcd59e029603b714022 Mon Sep 17 00:00:00 2001
From: 维石 <shixian.shi@alibaba-inc.com>
Date: 星期二, 11 六月 2024 11:52:26 +0800
Subject: [PATCH] english timestamp for valilla paraformer
---
funasr/auto/auto_model.py | 51 +++++++++++++++++++++++++++++++++++----------------
1 files changed, 35 insertions(+), 16 deletions(-)
diff --git a/funasr/auto/auto_model.py b/funasr/auto/auto_model.py
index 23d25c9..fb81608 100644
--- a/funasr/auto/auto_model.py
+++ b/funasr/auto/auto_model.py
@@ -19,6 +19,7 @@
from funasr.utils.load_utils import load_bytes
from funasr.download.file import download_from_url
from funasr.utils.timestamp_tools import timestamp_sentence
+from funasr.utils.timestamp_tools import timestamp_sentence_en
from funasr.download.download_from_hub import download_model
from funasr.utils.vad_utils import slice_padding_audio_samples
from funasr.utils.vad_utils import merge_vad
@@ -42,8 +43,9 @@
filelist = [".scp", ".txt", ".json", ".jsonl", ".text"]
chars = string.ascii_letters + string.digits
- if isinstance(data_in, str) and data_in.startswith("http://"): # url
- data_in = download_from_url(data_in)
+ if isinstance(data_in, str):
+ if data_in.startswith("http://") or data_in.startswith("https://"): # url
+ data_in = download_from_url(data_in)
if isinstance(data_in, str) and os.path.exists(
data_in
@@ -284,7 +286,7 @@
with torch.no_grad():
res = model.inference(**batch, **kwargs)
if isinstance(res, (list, tuple)):
- results = res[0]
+ results = res[0] if len(res) > 0 else [{"text": ""}]
meta_data = res[1] if len(res) > 1 else {}
time2 = time.perf_counter()
@@ -320,7 +322,7 @@
input, input_len=input_len, model=self.vad_model, kwargs=self.vad_kwargs, **cfg
)
end_vad = time.time()
-
+
# FIX(gcf): concat the vad clips for sense vocie model for better aed
if kwargs.get("merge_vad", False):
for i in range(len(res)):
@@ -358,6 +360,7 @@
results_sorted = []
if not len(sorted_data):
+ results_ret_list.append({"key": key, "text": "", "timestamp": []})
logging.info("decoding, utt: {}, empty speech".format(key))
continue
@@ -511,24 +514,40 @@
and 'iic/speech_seaco_paraformer_large_asr_nat-zh-cn-16k-common-vocab8404-pytorch'\
can predict timestamp, and speaker diarization relies on timestamps."
)
- sentence_list = timestamp_sentence(
- punc_res[0]["punc_array"],
- result["timestamp"],
- raw_text,
- return_raw_text=return_raw_text,
- )
+ if kwargs.get("en_post_proc", False):
+ sentence_list = timestamp_sentence_en(
+ punc_res[0]["punc_array"],
+ result["timestamp"],
+ raw_text,
+ return_raw_text=return_raw_text,
+ )
+ else:
+ sentence_list = timestamp_sentence(
+ punc_res[0]["punc_array"],
+ result["timestamp"],
+ raw_text,
+ return_raw_text=return_raw_text,
+ )
distribute_spk(sentence_list, sv_output)
result["sentence_info"] = sentence_list
elif kwargs.get("sentence_timestamp", False):
if not len(result["text"].strip()):
sentence_list = []
else:
- sentence_list = timestamp_sentence(
- punc_res[0]["punc_array"],
- result["timestamp"],
- raw_text,
- return_raw_text=return_raw_text,
- )
+ if kwargs.get("en_post_proc", False):
+ sentence_list = timestamp_sentence_en(
+ punc_res[0]["punc_array"],
+ result["timestamp"],
+ raw_text,
+ return_raw_text=return_raw_text,
+ )
+ else:
+ sentence_list = timestamp_sentence(
+ punc_res[0]["punc_array"],
+ result["timestamp"],
+ raw_text,
+ return_raw_text=return_raw_text,
+ )
result["sentence_info"] = sentence_list
if "spk_embedding" in result:
del result["spk_embedding"]
--
Gitblit v1.9.1