Organization of Pipelines
Preprocessing
database = preprocessing(orignal_database)
Dataloader
speech_sources, noise_source =
source_loader = Dataloader(dataset, parallel_preload, num_sources, shuffle, sample_rate, include_vad, max_norm use_cache, cache_args={max_cache, cache}, label="")
source_loader.__iter__
source_loader.__next__
# 传统的 clean dataset
preprocessing.py => clean_dataset
# database => Dataloader => return Sources
# __iter__，__next__ 等等应该如何设计，如果遍历，那么就顺序，否则就是随机生成？或者给一个专门的接口吧
# How to select sources from the different speakers?
source_loader = Dataloader(dataset, parallel_preload, num_sources, shuffle, sample_rate, include_vad, max_norm use_cache, cache_args={max_cache, cache}, label="")
noise_loader = Sourceloader(...)
rir_loader = ...
mic = Mic(array_arch)
room_loader = Roomloader(roomsize, min_allowable_distance_to_wall, angle_dist, use_elevation)
rir_simulate = RIRSimulator(backend="pyroom")
traj_generator = Trajectory_generator(min_allowable_distance_to_wall, use_elevation, use_azimuth, additional databases, types=[""])
# 传统的 clean piplines
# 将 pipeline 中的复杂度降低，这样你可以走多个 pipelines
for (sources:list[Source], noise: Source, room: Room, mic) in sources_loader:
noise: Source = next(noise_loader) # 或者 noise_loader.sample(外部参数)，实际上就是给一个 random seed 而已，哈哈
room: Room = room_generator.create() # 包含房间，墙，以及相关的系数
sources_traj = Trajectory_generator.generate(num_spk, sources: 时长) # (with random type)
rirs = rir_simulate.simulate_rirs(sources, mic, room, rir_interval) # 将 room 与 traj 等信息全部给到仿真器
# Source has y_rir, y_rir_early, ...
mix_source = mix(sources, noise, overlap_rate, target_loudness_level, floating_range_target loudness_level, mix_mode)
# steps，没有什么是可以替代思考的，思考就像跑步，不思考就会越来越慢
# 0. normalize sources gains.
# 1. source.gains = mixer.sample_source_gains(num_sources) # 因为要控制总体平均，所以需要使用类，而不是函数, (-5, -3, ...)
# 2. noise.gains = mixer.sample_noise_grans(1) # 也要控制总体
# 4. apply gains
#
# for-loop mixture, sources, noise
filter_options(output_options):
...
name = "source_4_1_noise_2_2_"