From 030043f768fa82c73e5decdf95f1016bf49b962a Mon Sep 17 00:00:00 2001
From: zhifu gao <zhifu.gzf@alibaba-inc.com>
Date: 星期四, 13 四月 2023 10:05:16 +0800
Subject: [PATCH] Merge pull request #341 from alibaba-damo-academy/dev_zly2
---
funasr/runtime/python/onnxruntime/funasr_onnx/vad_bin.py | 73 +++++++++++++++++++++---------------
1 files changed, 42 insertions(+), 31 deletions(-)
diff --git a/funasr/runtime/python/onnxruntime/funasr_onnx/vad_bin.py b/funasr/runtime/python/onnxruntime/funasr_onnx/vad_bin.py
index 221867d..5ad4266 100644
--- a/funasr/runtime/python/onnxruntime/funasr_onnx/vad_bin.py
+++ b/funasr/runtime/python/onnxruntime/funasr_onnx/vad_bin.py
@@ -59,37 +59,48 @@
def __call__(self, audio_in: Union[str, np.ndarray, List[str]], **kwargs) -> List:
- # waveform_list = self.load_data(audio_in, self.frontend.opts.frame_opts.samp_freq)
-
- param_dict = kwargs.get('param_dict', dict())
- is_final = param_dict.get('is_final', False)
- audio_in_cache = param_dict.get('audio_in_cache', None)
- audio_in_cum = audio_in
- if audio_in_cache is not None:
- audio_in_cum = np.concatenate((audio_in_cache, audio_in_cum))
- param_dict['audio_in_cache'] = audio_in_cum
- feats, feats_len = self.extract_feat([audio_in_cum])
-
- in_cache = param_dict.get('in_cache', list())
- in_cache = self.prepare_cache(in_cache)
- beg_idx = param_dict.get('beg_idx',0)
- feats = feats[:, beg_idx:beg_idx+8, :]
- param_dict['beg_idx'] = beg_idx + feats.shape[1]
- try:
- inputs = [feats]
- inputs.extend(in_cache)
- scores, out_caches = self.infer(inputs)
- param_dict['in_cache'] = out_caches
- segments = self.vad_scorer(scores, audio_in[None, :], is_final=is_final, max_end_sil=self.max_end_sil)
- # print(segments)
- if len(segments) == 1 and segments[0][0][1] != -1:
- self.frontend.reset_status()
+ waveform_list = self.load_data(audio_in, self.frontend.opts.frame_opts.samp_freq)
+ waveform_nums = len(waveform_list)
+ is_final = kwargs.get('kwargs', False)
+
+ segments = [[]] * self.batch_size
+ for beg_idx in range(0, waveform_nums, self.batch_size):
-
- except ONNXRuntimeError:
- logging.warning(traceback.format_exc())
- logging.warning("input wav is silence or noise")
- segments = []
+ end_idx = min(waveform_nums, beg_idx + self.batch_size)
+ waveform = waveform_list[beg_idx:end_idx]
+ feats, feats_len = self.extract_feat(waveform)
+ waveform = np.array(waveform)
+ param_dict = kwargs.get('param_dict', dict())
+ in_cache = param_dict.get('in_cache', list())
+ in_cache = self.prepare_cache(in_cache)
+ try:
+ t_offset = 0
+ step = int(min(feats_len.max(), 6000))
+ for t_offset in range(0, int(feats_len), min(step, feats_len - t_offset)):
+ if t_offset + step >= feats_len - 1:
+ step = feats_len - t_offset
+ is_final = True
+ else:
+ is_final = False
+ feats_package = feats[:, t_offset:int(t_offset + step), :]
+ waveform_package = waveform[:, t_offset * 160:min(waveform.shape[-1], (int(t_offset + step) - 1) * 160 + 400)]
+
+ inputs = [feats_package]
+ # inputs = [feats]
+ inputs.extend(in_cache)
+ scores, out_caches = self.infer(inputs)
+ in_cache = out_caches
+ segments_part = self.vad_scorer(scores, waveform_package, is_final=is_final, max_end_sil=self.max_end_sil, online=False)
+ # segments = self.vad_scorer(scores, waveform[0][None, :], is_final=is_final, max_end_sil=self.max_end_sil)
+
+ if segments_part:
+ for batch_num in range(0, self.batch_size):
+ segments[batch_num] += segments_part[batch_num]
+
+ except ONNXRuntimeError:
+ # logging.warning(traceback.format_exc())
+ logging.warning("input wav is silence or noise")
+ segments = ''
return segments
@@ -140,4 +151,4 @@
outputs = self.ort_infer(feats)
scores, out_caches = outputs[0], outputs[1:]
return scores, out_caches
-
\ No newline at end of file
+
--
Gitblit v1.9.1