From 1028a8a036cabd6091fc1a040bbddd565fd3e911 Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期三, 10 一月 2024 17:42:53 +0800
Subject: [PATCH] funasr1.0 paraformer_streaming WavFrontendOnline
---
funasr/bin/inference.py | 37 +++++++++++++++++++++----------------
1 files changed, 21 insertions(+), 16 deletions(-)
diff --git a/funasr/bin/inference.py b/funasr/bin/inference.py
index 1fac92e..c4ff69b 100644
--- a/funasr/bin/inference.py
+++ b/funasr/bin/inference.py
@@ -17,11 +17,12 @@
import string
from funasr.register import tables
-from funasr.utils.load_utils import load_audio_and_text_image_video, extract_fbank
+from funasr.utils.load_utils import load_audio_text_image_video, extract_fbank
from funasr.utils.vad_utils import slice_padding_audio_samples
from funasr.utils.timestamp_tools import time_stamp_sentence
+from funasr.download.file import download_from_url
-def build_iter_for_infer(data_in, input_len=None, data_type=None, key=None):
+def prepare_data_iterator(data_in, input_len=None, data_type=None, key=None):
"""
:param input:
@@ -35,7 +36,8 @@
filelist = [".scp", ".txt", ".json", ".jsonl"]
chars = string.ascii_letters + string.digits
-
+ if isinstance(data_in, str) and data_in.startswith('http'): # url
+ data_in = download_from_url(data_in)
if isinstance(data_in, str) and os.path.exists(data_in): # wav_path; filelist: wav.scp, file.jsonl;text.txt;
_, file_extension = os.path.splitext(data_in)
file_extension = file_extension.lower()
@@ -59,16 +61,16 @@
data_list = [data_in]
key_list = [key]
elif isinstance(data_in, (list, tuple)):
- if data_type is not None and isinstance(data_type, (list, tuple)):
+ if data_type is not None and isinstance(data_type, (list, tuple)): # mutiple inputs
data_list_tmp = []
for data_in_i, data_type_i in zip(data_in, data_type):
- key_list, data_list_i = build_iter_for_infer(data_in=data_in_i, data_type=data_type_i)
+ key_list, data_list_i = prepare_data_iterator(data_in=data_in_i, data_type=data_type_i)
data_list_tmp.append(data_list_i)
data_list = []
for item in zip(*data_list_tmp):
data_list.append(item)
else:
- # [audio sample point, fbank]
+ # [audio sample point, fbank, text]
data_list = data_in
key_list = ["rand_key_" + ''.join(random.choice(chars) for _ in range(13)) for _ in range(len(data_in))]
else: # raw text; audio sample point, fbank; bytes
@@ -198,13 +200,12 @@
kwargs = self.kwargs if kwargs is None else kwargs
kwargs.update(cfg)
model = self.model if model is None else model
-
- data_type = kwargs.get("data_type", "sound")
+
batch_size = kwargs.get("batch_size", 1)
# if kwargs.get("device", "cpu") == "cpu":
# batch_size = 1
- key_list, data_list = build_iter_for_infer(input, input_len=input_len, data_type=data_type, key=key)
+ key_list, data_list = prepare_data_iterator(input, input_len=input_len, data_type=kwargs.get("data_type", None), key=key)
speed_stats = {}
asr_result_list = []
@@ -222,7 +223,8 @@
batch["data_lengths"] = input_len
time1 = time.perf_counter()
- results, meta_data = model.generate(**batch, **kwargs)
+ with torch.no_grad():
+ results, meta_data = model.generate(**batch, **kwargs)
time2 = time.perf_counter()
asr_result_list.extend(results)
@@ -267,8 +269,8 @@
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
- data_type = kwargs.get("data_type", "sound")
- key_list, data_list = build_iter_for_infer(input, input_len=input_len, data_type=data_type)
+
+ 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 = 0.0
@@ -278,7 +280,7 @@
key = res[i]["key"]
vadsegments = res[i]["value"]
input_i = data_list[i]
- speech = load_audio_and_text_image_video(input_i, fs=kwargs["frontend"].fs, audio_fs=kwargs.get("fs", 16000))
+ speech = load_audio_text_image_video(input_i, fs=kwargs["frontend"].fs, audio_fs=kwargs.get("fs", 16000))
speech_lengths = len(speech)
n = len(vadsegments)
data_with_index = [(vadsegments[i], i) for i in range(n)]
@@ -389,7 +391,10 @@
frontend = frontend_class(**kwargs["frontend_conf"])
self.frontend = frontend
+ if "frontend" in kwargs:
+ del kwargs["frontend"]
self.kwargs = kwargs
+
def __call__(self, input, input_len=None, kwargs=None, **cfg):
@@ -397,7 +402,7 @@
kwargs.update(cfg)
- key_list, data_list = build_iter_for_infer(input, input_len=input_len)
+ key_list, data_list = prepare_data_iterator(input, input_len=input_len)
batch_size = kwargs.get("batch_size", 1)
device = kwargs.get("device", "cpu")
if device == "cpu":
@@ -417,11 +422,11 @@
# extract fbank feats
time1 = time.perf_counter()
- audio_sample_list = load_audio_and_text_image_video(data_batch, fs=self.frontend.fs, audio_fs=kwargs.get("fs", 16000))
+ audio_sample_list = load_audio_text_image_video(data_batch, fs=self.frontend.fs, audio_fs=kwargs.get("fs", 16000))
time2 = time.perf_counter()
meta_data["load_data"] = f"{time2 - time1:0.3f}"
speech, speech_lengths = extract_fbank(audio_sample_list, data_type=kwargs.get("data_type", "sound"),
- frontend=self.frontend)
+ frontend=self.frontend, **kwargs)
time3 = time.perf_counter()
meta_data["extract_feat"] = f"{time3 - time2:0.3f}"
meta_data["batch_data_time"] = speech_lengths.sum().item() * self.frontend.frame_shift * self.frontend.lfr_n / 1000
--
Gitblit v1.9.1