From fa56f36921c6bcb4608a28ab76686822033b728e Mon Sep 17 00:00:00 2001
From: shixian.shi <shixian.shi@alibaba-inc.com>
Date: 星期四, 25 一月 2024 18:48:48 +0800
Subject: [PATCH] Update demo.py
---
funasr/auto/auto_model.py | 31 ++++++++++++++++---------------
1 files changed, 16 insertions(+), 15 deletions(-)
diff --git a/funasr/auto/auto_model.py b/funasr/auto/auto_model.py
index 0641f06..4d0f302 100644
--- a/funasr/auto/auto_model.py
+++ b/funasr/auto/auto_model.py
@@ -123,7 +123,6 @@
self.preset_spk_num = kwargs.get("preset_spk_num", None)
if self.preset_spk_num:
logging.warning("Using preset speaker number: {}".format(self.preset_spk_num))
- logging.warning("Many to print when using speaker model...")
self.kwargs = kwargs
self.model = model
@@ -146,7 +145,7 @@
set_all_random_seed(kwargs.get("seed", 0))
device = kwargs.get("device", "cuda")
- if not torch.cuda.is_available() or kwargs.get("ngpu", 0):
+ if not torch.cuda.is_available() or kwargs.get("ngpu", 0) == 0:
device = "cpu"
kwargs["batch_size"] = 1
kwargs["device"] = device
@@ -224,7 +223,7 @@
asr_result_list = []
num_samples = len(data_list)
disable_pbar = kwargs.get("disable_pbar", False)
- pbar = tqdm(colour="blue", total=num_samples+1, dynamic_ncols=True) if not disable_pbar else None
+ pbar = tqdm(colour="blue", total=num_samples, dynamic_ncols=True) if not disable_pbar else None
time_speech_total = 0.0
time_escape_total = 0.0
for beg_idx in range(0, num_samples, batch_size):
@@ -311,7 +310,7 @@
batch_size_ms_cum = 0
beg_idx = 0
beg_asr_total = time.time()
- time_speech_total_per_sample = speech_lengths/16000
+ time_speech_total_per_sample = speech_lengths/16000 + 1e-6
time_speech_total_all_samples += time_speech_total_per_sample
pbar_sample = tqdm(colour="blue", total=n + 1, dynamic_ncols=True)
@@ -329,8 +328,6 @@
speech_j, speech_lengths_j = slice_padding_audio_samples(speech, speech_lengths, sorted_data[beg_idx:end_idx])
results = self.inference(speech_j, input_len=None, model=model, kwargs=kwargs, disable_pbar=True, **cfg)
if self.spk_model is not None:
-
-
# compose vad segments: [[start_time_sec, end_time_sec, speech], [...]]
for _b in range(len(speech_j)):
vad_segments = [[sorted_data[beg_idx:end_idx][_b][0][0]/1000.0,
@@ -345,16 +342,14 @@
if len(results) < 1:
continue
results_sorted.extend(results)
-
-
end_asr_total = time.time()
time_escape_total_per_sample = end_asr_total - beg_asr_total
+ pbar_sample.update(1)
pbar_sample.set_description(f"rtf_avg_per_sample: {time_escape_total_per_sample / time_speech_total_per_sample:0.3f}, "
f"time_speech_total_per_sample: {time_speech_total_per_sample: 0.3f}, "
f"time_escape_total_per_sample: {time_escape_total_per_sample:0.3f}")
-
restored_data = [0] * n
for j in range(n):
index = sorted_data[j][1]
@@ -377,7 +372,7 @@
result[k] = restored_data[j][k]
else:
result[k] = torch.cat([result[k], restored_data[j][k]], dim=0)
- elif k == 'text':
+ elif 'text' in k:
if k not in result:
result[k] = restored_data[j][k]
else:
@@ -392,8 +387,9 @@
if self.punc_model is not None:
self.punc_kwargs.update(cfg)
punc_res = self.inference(result["text"], model=self.punc_model, kwargs=self.punc_kwargs, **cfg)
+ import copy; raw_text = copy.copy(result["text"])
result["text"] = punc_res[0]["text"]
-
+
# speaker embedding cluster after resorted
if self.spk_model is not None:
all_segments = sorted(all_segments, key=lambda x: x[0])
@@ -401,19 +397,24 @@
labels = self.cb_model(spk_embedding.cpu(), oracle_num=self.preset_spk_num)
del result['spk_embedding']
sv_output = postprocess(all_segments, None, labels, spk_embedding.cpu())
- if self.spk_mode == 'vad_segment':
+ if self.spk_mode == 'vad_segment': # recover sentence_list
sentence_list = []
for res, vadsegment in zip(restored_data, vadsegments):
sentence_list.append({"start": vadsegment[0],\
"end": vadsegment[1],
- "sentence": res['text'],
+ "sentence": res['raw_text'],
"timestamp": res['timestamp']})
- else: # punc_segment
+ elif self.spk_mode == 'punc_segment':
sentence_list = timestamp_sentence(punc_res[0]['punc_array'], \
result['timestamp'], \
- result['text'])
+ result['raw_text'])
distribute_spk(sentence_list, sv_output)
result['sentence_info'] = sentence_list
+ elif kwargs.get("sentence_timestamp", False):
+ sentence_list = timestamp_sentence(punc_res[0]['punc_array'], \
+ result['timestamp'], \
+ result['raw_text'])
+ result['sentence_info'] = sentence_list
result["key"] = key
results_ret_list.append(result)
--
Gitblit v1.9.1