add corrected applicator to linear alignment (#407)

* add corrected applicator to linear alignment

* add sis hash exclude

* revert

* change import and rerevert

* more detailed import

* exclude test

* possible fix

* black

* more

* add doc string

* removed -> py3.8

am/config.py

@@ -1,14 +1,14 @@
 __all__ = ["acoustic_model_config", "get_align_config_and_crp_for_corrected_applicator"]
 
-import copy
 from dataclasses import dataclass
 from typing import List, Literal, Tuple, Union, Optional
 
-import i6_core.rasr as rasr
-
 from sisyphus import tk
 
+import i6_core.rasr as rasr
+
 TdpType = Union[float, Literal["infinity"]]
+TyingType = Literal["global", "global-and-nonword"]
 
 
 @dataclass
@@ -28,7 +28,7 @@ def acoustic_model_config(
     tdp_scale: float = 1.0,
     tdp_transition: Union[TdpValues, Tuple[TdpType, ...]] = (3.0, 0.0, 3.0, 2.0),
     tdp_silence: Union[TdpValues, Tuple[TdpType, ...]] = (0.0, 3.0, "infinity", 6.0),
-    tying_type: Literal["global", "global-and-nonword"] = "global",
+    tying_type: TyingType = "global",
     nonword_phones: Union[str, List[str]] = "",
     tdp_nonword: Union[TdpValues, Tuple[TdpType, ...]] = (0.0, 3.0, "infinity", 6.0),
     state_tying_file: Optional[tk.Path] = None,
@@ -111,9 +111,10 @@ def get_align_config_and_crp_for_corrected_applicator(
     crp: rasr.CommonRasrParameters, exit_penalty: float = 0.0
 ) -> [rasr.CommonRasrParameters, rasr.RasrConfig]:
     """
-    Set the correct type of applicator, default is "legacy". Moreover, set exit penalities to zero
+    Set the correct type of applicator, default is "legacy". Moreover, set exit penalties to zero
     For a given word sequence the exit penalty is constant with respect to the max/sum
-    :param config:
+    :param crp: common rasr parameters
+    :param exit_penalty: exit penalty for speech phonemes, silence, and non word phonemes
     :return:
     """
 

mm/mixtures.py

@@ -21,6 +21,8 @@
 import i6_core.rasr as rasr
 import i6_core.util as util
 
+from i6_core.am.config import get_align_config_and_crp_for_corrected_applicator
+
 
 class MergeMixturesJob(rasr.RasrCommand, Job):
     def __init__(
@@ -147,6 +149,8 @@ def hash(cls, kwargs):
 
 
 class LinearAlignmentJob(MergeMixturesJob):
+    __sis_hash_exclude__ = {"use_corrected_applicator": False}
+
     def __init__(
         self,
         crp,
@@ -158,6 +162,8 @@ def __init__(
         minimum_speech_proportion=0.7,
         save_alignment=False,
         keep_accumulators=False,
+        use_corrected_applicator: bool = False,
+        exit_penalty_for_corrected_applicator: float = 0.0,
         extra_merge_args=None,
         extra_config=None,
         extra_post_config=None,
@@ -247,10 +253,19 @@ def create_config(
         penalty,
         minimum_speech_proportion,
         save_alignment,
+        use_corrected_applicator,
+        exit_penalty_for_corrected_applicator,
         extra_config,
         extra_post_config,
         **kwargs,
     ):
+        if use_corrected_applicator:
+            crp, _extra_config = get_align_config_and_crp_for_corrected_applicator(
+                crp=crp,
+                exit_penalty=exit_penalty_for_corrected_applicator,
+            )
+            extra_config._update(_extra_config)
+
         segmentation_flow = cls.create_flow(feature_energy_flow, save_alignment)
         mapping = {
             "corpus": "acoustic-model-trainer.corpus",