From 4ba1011b42e041ee1d71448eefd7ef2e7bd61bb6 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期五, 31 三月 2023 15:31:26 +0800
Subject: [PATCH] export
---
funasr/runtime/python/onnxruntime/funasr_onnx/vad_bin.py | 71 ++++++++++++++++++++---------------
1 files changed, 40 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 533b4b7..221867d 100644
--- a/funasr/runtime/python/onnxruntime/funasr_onnx/vad_bin.py
+++ b/funasr/runtime/python/onnxruntime/funasr_onnx/vad_bin.py
@@ -23,7 +23,7 @@
device_id: Union[str, int] = "-1",
quantize: bool = False,
intra_op_num_threads: int = 4,
- max_end_sil: int = 800,
+ max_end_sil: int = None,
):
if not Path(model_dir).exists():
@@ -43,46 +43,55 @@
self.ort_infer = OrtInferSession(model_file, device_id, intra_op_num_threads=intra_op_num_threads)
self.batch_size = batch_size
self.vad_scorer = E2EVadModel(config["vad_post_conf"])
- self.max_end_sil = max_end_sil
+ self.max_end_sil = max_end_sil if max_end_sil is not None else config["vad_post_conf"]["max_end_silence_time"]
+ self.encoder_conf = config["encoder_conf"]
def prepare_cache(self, in_cache: list = []):
if len(in_cache) > 0:
return in_cache
-
- for i in range(4):
- cache = np.random.rand(1, 128, 19, 1).astype(np.float32)
+ fsmn_layers = self.encoder_conf["fsmn_layers"]
+ proj_dim = self.encoder_conf["proj_dim"]
+ lorder = self.encoder_conf["lorder"]
+ for i in range(fsmn_layers):
+ cache = np.zeros((1, proj_dim, lorder-1, 1)).astype(np.float32)
in_cache.append(cache)
return in_cache
- def __call__(self, wav_content: Union[str, np.ndarray, List[str]], **kwargs) -> List:
- waveform_list = self.load_data(wav_content, self.frontend.opts.frame_opts.samp_freq)
- waveform_nums = len(waveform_list)
- is_final = kwargs.get('kwargs', False)
-
- asr_res = []
- for beg_idx in range(0, waveform_nums, self.batch_size):
+ 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()
- end_idx = min(waveform_nums, beg_idx + self.batch_size)
- waveform = waveform_list[beg_idx:end_idx]
- feats, feats_len = self.extract_feat(waveform)
- param_dict = kwargs.get('param_dict', dict())
- in_cache = param_dict.get('cache', list())
- in_cache = self.prepare_cache(in_cache)
- try:
- inputs = [feats]
- inputs.extend(in_cache)
- scores, out_caches = self.infer(inputs)
- param_dict['cache'] = out_caches
- segments = self.vad_scorer(scores, waveform[0][None, :], is_final=is_final, max_end_sil=self.max_end_sil)
-
- except ONNXRuntimeError:
- # logging.warning(traceback.format_exc())
- logging.warning("input wav is silence or noise")
- segments = ''
- asr_res.append(segments)
+
+ except ONNXRuntimeError:
+ logging.warning(traceback.format_exc())
+ logging.warning("input wav is silence or noise")
+ segments = []
- return asr_res
+ return segments
def load_data(self,
wav_content: Union[str, np.ndarray, List[str]], fs: int = None) -> List:
--
Gitblit v1.9.1