python/FunASR-XL.git

			@@ -19,7 +19,8 @@
			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.download.download_from_hub import download_model
			from funasr.utils.timestamp_tools import timestamp_sentence_en
			from funasr.download.download_model_from_hub import download_model
			from funasr.utils.vad_utils import slice_padding_audio_samples
			from funasr.utils.vad_utils import merge_vad
			from funasr.utils.load_utils import load_audio_text_image_video
			@@ -91,7 +92,8 @@
			if isinstance(data_i, str) and os.path.exists(data_i):
			key = misc.extract_filename_without_extension(data_i)
			else:
			key = "rand_key_" + "".join(random.choice(chars) for _ in range(13))
			if key is None:
			key = "rand_key_" + "".join(random.choice(chars) for _ in range(13))
			key_list.append(key)

			else: # raw text; audio sample point, fbank; bytes
			@@ -109,11 +111,15 @@

			def __init__(self, **kwargs):

			try:
			from funasr.utils.version_checker import check_for_update

			check_for_update()
			except:
			pass

			log_level = getattr(logging, kwargs.get("log_level", "INFO").upper())
			logging.basicConfig(level=log_level)

			if not kwargs.get("disable_log", True):
			tables.print()

			model, kwargs = self.build_model(**kwargs)

			@@ -162,7 +168,8 @@
			self.spk_kwargs = spk_kwargs
			self.model_path = kwargs.get("model_path")

			def build_model(self, **kwargs):
			@staticmethod
			def build_model(**kwargs):
			assert "model" in kwargs
			if "model_conf" not in kwargs:
			logging.info("download models from model hub: {}".format(kwargs.get("hub", "ms")))
			@@ -208,11 +215,11 @@
			kwargs["frontend"] = frontend
			# build model
			model_class = tables.model_classes.get(kwargs["model"])
			assert model_class is not None, f'{kwargs["model"]} is not registered'
			model_conf = {}
			deep_update(model_conf, kwargs.get("model_conf", {}))
			deep_update(model_conf, kwargs)
			model = model_class(**model_conf, vocab_size=vocab_size)
			model.to(device)

			# init_param
			init_param = kwargs.get("init_param", None)
			@@ -233,6 +240,13 @@
			# fp16
			if kwargs.get("fp16", False):
			model.to(torch.float16)
			elif kwargs.get("bf16", False):
			model.to(torch.bfloat16)
			model.to(device)

			if not kwargs.get("disable_log", True):
			tables.print()

			return model, kwargs

			def __call__(self, args, *cfg):
			@@ -301,7 +315,7 @@
			speed_stats["rtf"] = f"{(time_escape) / batch_data_time:0.3f}"
			description = f"{speed_stats}, "
			if pbar:
			pbar.update(1)
			pbar.update(end_idx - beg_idx)
			pbar.set_description(description)
			time_speech_total += batch_data_time
			time_escape_total += time_escape
			@@ -427,6 +441,10 @@
			# 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}")

			if len(results_sorted) != n:
			results_ret_list.append({"key": key, "text": "", "timestamp": []})
			logging.info("decoding, utt: {}, empty result".format(key))
			continue
			restored_data = [0] * n
			for j in range(n):
			index = sorted_data[j][1]
			@@ -460,23 +478,20 @@
			else:
			result[k] += restored_data[j][k]

			if not len(result["text"].strip()):
			continue
			return_raw_text = kwargs.get("return_raw_text", False)
			# step.3 compute punc model
			raw_text = None
			if self.punc_model is not None:
			if not len(result["text"].strip()):
			if return_raw_text:
			result["raw_text"] = ""
			else:
			deep_update(self.punc_kwargs, 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
			deep_update(self.punc_kwargs, 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"]

			# speaker embedding cluster after resorted
			if self.spk_model is not None and kwargs.get("return_spk_res", True):
			@@ -513,24 +528,40 @@
			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"]
			@@ -582,12 +613,6 @@
			)

			with torch.no_grad():

			if type == "onnx":
			export_dir = export_utils.export_onnx(model=model, data_in=data_list, **kwargs)
			else:
			export_dir = export_utils.export_torchscripts(
			model=model, data_in=data_list, **kwargs
			)
			export_dir = export_utils.export(model=model, data_in=data_list, **kwargs)

			return export_dir