python/FunASR-XL.git

			@@ -219,7 +219,7 @@
			speed_stats = {}
			asr_result_list = []
			num_samples = len(data_list)
			disable_pbar = kwargs.get("disable_pbar", False)
			disable_pbar = self.kwargs.get("disable_pbar", False)
			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
			@@ -231,12 +231,12 @@
			if (end_idx - beg_idx) == 1 and kwargs.get("data_type", None) == "fbank": # fbank
			batch["data_in"] = data_batch[0]
			batch["data_lengths"] = input_len


			time1 = time.perf_counter()
			with torch.no_grad():
			results, meta_data = model.inference(batch, kwargs)
			time2 = time.perf_counter()


			asr_result_list.extend(results)

			# batch_data_time = time_per_frame_s * data_batch_i["speech_lengths"].sum().item()
			@@ -261,31 +261,29 @@
			pbar.set_description(f"rtf_avg: {time_escape_total/time_speech_total:0.3f}")
			torch.cuda.empty_cache()
			return asr_result_list


			def inference_with_vad(self, input, input_len=None, **cfg):

			kwargs = self.kwargs
			# step.1: compute the vad model
			self.vad_kwargs.update(cfg)
			beg_vad = time.time()
			res = self.inference(input, input_len=input_len, model=self.vad_model, kwargs=self.vad_kwargs, **cfg)
			end_vad = time.time()
			print(f"time cost vad: {end_vad - beg_vad:0.3f}")


			# step.2 compute asr model
			model = self.model
			kwargs = self.kwargs
			kwargs.update(cfg)
			batch_size = int(kwargs.get("batch_size_s", 300))*1000
			batch_size_threshold_ms = int(kwargs.get("batch_size_threshold_s", 60))*1000
			kwargs["batch_size"] = batch_size


			key_list, data_list = prepare_data_iterator(input, input_len=input_len, data_type=kwargs.get("data_type", None))
			results_ret_list = []
			time_speech_total_all_samples = 1e-6

			beg_total = time.time()
			pbar_total = tqdm(colour="red", total=len(res), dynamic_ncols=True)
			pbar_total = tqdm(colour="red", total=len(res), dynamic_ncols=True) if not kwargs.get("disable_pbar", False) else None
			for i in range(len(res)):
			key = res[i]["key"]
			vadsegments = res[i]["value"]
			@@ -296,14 +294,14 @@
			data_with_index = [(vadsegments[i], i) for i in range(n)]
			sorted_data = sorted(data_with_index, key=lambda x: x[0][1] - x[0][0])
			results_sorted = []


			if not len(sorted_data):
			logging.info("decoding, utt: {}, empty speech".format(key))
			continue

			if len(sorted_data) > 0 and len(sorted_data[0]) > 0:
			batch_size = max(batch_size, sorted_data[0][0][1] - sorted_data[0][0][0])


			batch_size_ms_cum = 0
			beg_idx = 0
			beg_asr_total = time.time()
			@@ -322,8 +320,8 @@
			continue
			batch_size_ms_cum = 0
			end_idx = j + 1
			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)
			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, **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)):
			@@ -333,26 +331,26 @@
			segments = sv_chunk(vad_segments)
			all_segments.extend(segments)
			speech_b = [i[2] for i in segments]
			spk_res = self.inference(speech_b, input_len=None, model=self.spk_model, kwargs=kwargs, disable_pbar=True, **cfg)
			spk_res = self.inference(speech_b, input_len=None, model=self.spk_model, kwargs=kwargs, **cfg)
			results[_b]['spk_embedding'] = spk_res[0]['spk_embedding']
			beg_idx = end_idx
			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]
			restored_data[index] = results_sorted[j]
			result = {}


			# results combine for texts, timestamps, speaker embeddings and others
			# TODO: rewrite for clean code
			for j in range(n):
			@@ -379,18 +377,18 @@
			result[k] = restored_data[j][k]
			else:
			result[k] += restored_data[j][k]

			return_raw_text = kwargs.get('return_raw_text', False)

			return_raw_text = kwargs.get('return_raw_text', False)
			# step.3 compute punc model
			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, disable_pbar=True, **cfg)
			punc_res = self.inference(result["text"], model=self.punc_model, kwargs=self.punc_kwargs, **cfg)
			raw_text = copy.copy(result["text"])
			if return_raw_text: result['raw_text'] = raw_text
			result["text"] = punc_res[0]["text"]
			else:
			raw_text = None


			# speaker embedding cluster after resorted
			if self.spk_model is not None and kwargs.get('return_spk_res', True):
			if raw_text is None:
			@@ -429,13 +427,14 @@
			return_raw_text=return_raw_text)
			result['sentence_info'] = sentence_list
			if "spk_embedding" in result: del result['spk_embedding']


			result["key"] = key
			results_ret_list.append(result)
			end_asr_total = time.time()
			time_escape_total_per_sample = end_asr_total - beg_asr_total
			pbar_total.update(1)
			pbar_total.set_description(f"rtf_avg: {time_escape_total_per_sample / time_speech_total_per_sample:0.3f}, "
			if pbar_total:
			pbar_total.update(1)
			pbar_total.set_description(f"rtf_avg: {time_escape_total_per_sample / time_speech_total_per_sample:0.3f}, "
			f"time_speech: {time_speech_total_per_sample: 0.3f}, "
			f"time_escape: {time_escape_total_per_sample:0.3f}")