Fix g2p_model_path and add tests (#939)

Author: mmcauliffe

.github/workflows/main.yml

@@ -14,33 +14,24 @@ concurrency:
 
 jobs:
   build:
-    strategy:
-      matrix:
-        include:
-          - os: ubuntu-latest
-            label: linux-64
-
-          #- os: macos-latest
-          #  label: osx-64
-          #  prefix: /Users/runner/miniconda3/envs/my-env
-
-          #- os: windows-latest
-          #  label: win-64
-          #  prefix: C:\Miniconda3\envs\my-env
-
-    name: ${{ matrix.label }}
-    runs-on: ${{ matrix.os }}
+    runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@main
         with:
           fetch-depth: 0
 
+      - name: Check disk usage
+        run: df -h
+
       - name: Install Conda environment with Micromamba
         uses: conda-incubator/setup-miniconda@v3
         with:
           environment-file: github_environment.yml
           miniforge-version: latest
 
+      - name: Check disk usage after environment install
+        run: df -h
+
       - name: Configure mfa
         shell: bash -l {0}
         run: python -m montreal_forced_aligner configure --disable_auto_server

github_environment.yml

@@ -1,8 +1,8 @@
 channels:
   - conda-forge
 dependencies:
-  - python=3.11
-  - numpy
+  - python=3.12
+  - numpy<2.1.0
   - librosa
   - pysoundfile
   - tqdm
@@ -42,15 +42,28 @@ dependencies:
   - kalpy>=0.8
   - pip
   - huggingface_hub
+  - spacy
+  - sudachipy
+  - sudachidict-core
+  - jamo
+  - pythainlp
+  - python-build
+  - twine
+  # speechbrain dependencies
+  - transformers>=4.48.0
+  - pandas<2.3.0
+  - sentencepiece
+  - pytorch=2.8.0
+  - torchaudio=2.8.0
+  # whisperx dependencies
+  - triton
+  - nltk
+  - av<16.0.0
+  - onnxruntime<1.20.0
   - pip:
-      - build
-      - twine
+      - speechbrain
+      - whisperx
       # Tokenization dependencies
       - python-mecab-ko
-      - jamo
-      - pythainlp
       - hanziconv
       - dragonmapper
-      - speechbrain
-      - transformers
-      - whisperx

montreal_forced_aligner/command_line/align_one.py

@@ -139,10 +139,7 @@ def align_one_cli(context, **kwargs) -> None:
         lexicon_compiler.phone_table = pywrapfst.SymbolTable.read_text(phones_path)
     else:
         lexicon_compiler.load_pronunciations(dictionary_path)
-        lexicon_compiler.fst.write(str(l_fst_path))
-        lexicon_compiler.align_fst.write(str(l_align_fst_path))
-        lexicon_compiler.word_table.write_text(words_path)
-        lexicon_compiler.phone_table.write_text(phones_path)
+        lexicon_compiler.create_fsts()
         lexicon_compiler.clear()
 
     if no_tokenization or acoustic_model.language is Language.unknown:
@@ -193,6 +190,11 @@ def align_one_cli(context, **kwargs) -> None:
             "boost_silence",
         ]
     }
+    if g2p_model is not None or not (l_fst_path.exists() and not config.CLEAN):
+        lexicon_compiler.fst.write(str(l_fst_path))
+        lexicon_compiler.align_fst.write(str(l_align_fst_path))
+        lexicon_compiler.word_table.write_text(words_path)
+        lexicon_compiler.phone_table.write_text(phones_path)
     kalpy_aligner = KalpyAligner(acoustic_model, lexicon_compiler, **align_options)
     for utt in utterances:
         utt.apply_cmvn(cmvn)

montreal_forced_aligner/data.py

@@ -1573,6 +1573,24 @@ class WordData:
     pronunciations: typing.Set[typing.Tuple[str, ...]]
 
 
+# noinspection PyUnresolvedReferences
+@dataclassy.dataclass(slots=True)
+class GeneratedPronunciation:
+    """
+    Data class for information about a pronunciation generated by G2P models
+
+    Parameters
+    ----------
+    pronunciation: str
+        String of phones separated by spaces
+    score: float
+        Log-likelihood score for the pronunciation
+    """
+
+    pronunciation: str
+    score: float
+
+
 # noinspection PyUnresolvedReferences
 @dataclassy.dataclass(slots=True)
 class NgramHistoryState: