Merge pull request #14 from EveryVoiceTTS/dev.ej/326

Several improvements to wav2vec2aligner system: early errors, testing, py 3.8 compatibility, etc

.coveragerc

@@ -0,0 +1,14 @@
+[run]
+source_pkgs = aligner
+omit =
+    *tmp*
+    */run_tests.py
+    */tests/*
+    */__main__.py
+
+[report]
+precision = 2
+exclude_lines =
+    pragma: no cover
+    if 0:
+    if __name__ == .__main__.:

.github/workflows/test.yml

@@ -13,19 +13,38 @@ jobs:
       run:
         shell: bash -l {0}
     steps:
-      - name: Checkout repository
-        uses: actions/checkout@v4
-      - name: Set up Python
-        uses: actions/setup-python@v5
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
         with:
           python-version: "3.8"
           cache: "pip"
+
+      - uses: FedericoCarboni/setup-ffmpeg@v2
+
       - name: Install dependencies and package
-        run: pip install -e . mypy
-      - name: Minimal test, --help should work
-        run: ctc-segmenter --help
-      - name: Code quality test, mypy should pass
+        run: pip install -e . mypy coverage
+
+      - name: Minimal code quality test, mypy should pass
         run: mypy aligner
+
+      - uses: actions/cache@v4
+        with:
+          path: /home/runner/.cache/torch
+          key: torch-cache
+
+      - name: Run unit tests
+        run: |
+          coverage run -m unittest discover aligner.tests -v
+          coverage xml
+
+      - name: Upload coverage report to Codecov
+        uses: codecov/codecov-action@v4
+        with:
+          token: ${{ secrets.CODECOV_TOKEN }}
+
+      - run: coverage report
+
       - name: Make sure the CLI stays fast
         id: cli-load-time
         run: |
@@ -44,6 +63,7 @@ jobs:
             echo "Please run 'PYTHONPROFILEIMPORTTIME=1 ctc-segmenter -h 2> importtime.txt; tuna importtime.txt' and tuck away expensive imports so that the CLI doesn't load them until it uses them."; \
             false; \
           fi
+
       - name: Report help speed in PR
         if: github.event_name == 'pull_request'
         uses: mshick/add-pr-comment@v2

.gitignore

@@ -1,4 +1,4 @@
 *.egg-info
 __pycache__
 *.wav
-*.TextGrid
+*.TextGrid.coverage

aligner/__main__.py

@@ -0,0 +1,3 @@
+from .cli import app
+
+app()

aligner/cli.py

@@ -134,7 +134,14 @@ def align_single(
     ),
     debug: bool = typer.Option(False, help="Print debug statements"),
 ):
-    print("loading model...")
+    # Do fast error checking before loading expensive dependencies
+    sentence_list = read_text(text_path)
+    if not sentence_list or not any(sentence_list):
+        raise typer.BadParameter(
+            f"TEXT_PATH file '{text_path}' is empty; it should contain sentences to align.",
+        )
+
+    print("loading pytorch...")
     import torch
     import torchaudio
 
@@ -155,7 +162,6 @@ def align_single(
         audio_path = Path(fn + f"-{sample_rate}-mono" + ext)
         torchaudio.save(str(audio_path), wav, sample_rate)
     print("processing text")
-    sentence_list = read_text(text_path)
     transducer = create_transducer("".join(sentence_list), labels, debug)
     text_hash = TextHash(sentence_list, transducer)
     print("performing alignment")
@@ -180,7 +186,3 @@ def align_single(
     tg_path = audio_path.with_suffix(".TextGrid")
     print(f"writing file to {tg_path}")
     tg.to_file(tg_path)
-
-
-if __name__ == "__main__":
-    align_single()

aligner/heavy.py

@@ -79,7 +79,7 @@ def compute_alignments(
             token_index += 1
         frames.append(Frame(token_index, i, score))
         prev_hyp = ali
-    words_to_match = [v | {"key": k} for k, v in transcript_hash.items() if "w" in k]
+    words_to_match = [{**v, "key": k} for k, v in transcript_hash.items() if "w" in k]
     i1, i2 = 0, 0
     segments = []
     while i1 < len(frames):

aligner/tests/test_cli.py

@@ -0,0 +1,114 @@
+"""
+Run wav2vec2aligner unit tests.
+How to run this test suite:
+If you installed wav2vec2aligner:
+    python -m unittest aligner.tests.test_cli
+If you installed everyvoice:
+    python -m unittest everyvoice.model.aligner.wav2vec2aligner.aligner.tests.test_cli
+"""
+
+import os
+import subprocess
+import tempfile
+from pathlib import Path
+from unittest import TestCase
+
+from typer.testing import CliRunner
+
+from ..classes import Segment
+from ..cli import app, complete_path
+
+
+class CLITest(TestCase):
+    def setUp(self) -> None:
+        self.runner = CliRunner()
+
+    def test_main_help(self):
+        for help in "-h", "--help":
+            with self.subTest(help=help):
+                result = self.runner.invoke(app, [help])
+                self.assertEqual(result.exit_code, 0)
+                self.assertIn("align", result.stdout)
+                self.assertIn("extract", result.stdout)
+
+    def test_sub_help(self):
+        for cmd in "align", "extract":
+            for help in "-h", "--help":
+                with self.subTest(cmd=cmd, help=help):
+                    result = self.runner.invoke(app, [cmd, help])
+                    self.assertEqual(result.exit_code, 0)
+                    self.assertIn("Usage:", result.stdout)
+                    self.assertIn(cmd, result.stdout)
+
+    def test_align_empty_file(self):
+        with self.subTest("empty file"):
+            result = self.runner.invoke(app, ["align", os.devnull, os.devnull])
+            self.assertNotEqual(result.exit_code, 0)
+            self.assertIn("is empty", result.stdout)
+
+        with self.subTest("file with only empty lines"):
+            with tempfile.TemporaryDirectory() as tmpdir:
+                textfile = os.path.join(tmpdir, "emptylines.txt")
+                with open(textfile, "w", encoding="utf8") as f:
+                    f.write("\n \n   \n")
+                result = self.runner.invoke(app, ["align", textfile, os.devnull])
+                self.assertNotEqual(result.exit_code, 0)
+                self.assertIn("is empty", result.stdout)
+
+    def fetch_ras_test_file(self, filename, outputdir):
+        from urllib.request import Request, urlopen
+
+        repo, path = "https://github.com/ReadAlongs/Studio/", "/test/data/"
+        request = Request(repo + "raw/refs/heads/main" + path + filename)
+        request.add_header("Referer", repo + "blob/main" + path + filename)
+        response = urlopen(request)
+        with open(os.path.join(outputdir, filename), "wb") as f:
+            f.write(response.read())
+
+    def test_align_something(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            tmppath = Path(tmpdir)
+            self.fetch_ras_test_file("ej-fra.txt", tmpdir)
+            txt = tmppath / "ej-fra.txt"
+            self.fetch_ras_test_file("ej-fra.m4a", tmpdir)
+            m4a = tmppath / "ej-fra.m4a"
+            wav = tmppath / "ej-fra.wav"
+            # Under most circumstances, align can take a .m4a input file, but not
+            # in CI. Since it's not a hard requirement, so just convert to .wav.
+            subprocess.run(["ffmpeg", "-i", m4a, wav], capture_output=True)
+            # os.system("ls -la " + tmpdir)
+            textgrid = tmppath / "ej-fra-16000.TextGrid"
+            wav_out = tmppath / "ej-fra-16000.wav"
+
+            with self.subTest("ctc-segmenter align"):
+                result = self.runner.invoke(app, ["align", str(txt), str(wav)])
+                if result.exit_code != 0:
+                    os.system("ls -la " + tmpdir)
+                    print(result.stdout)
+                self.assertEqual(result.exit_code, 0)
+                self.assertTrue(textgrid.exists())
+                self.assertTrue(wav_out.exists())
+
+            with self.subTest("ctc-segmenter extract"):
+                result = self.runner.invoke(
+                    app, ["extract", str(textgrid), str(wav_out), str(tmppath / "out")]
+                )
+                if result.exit_code != 0:
+                    print(result.stdout)
+                self.assertEqual(result.exit_code, 0)
+                self.assertTrue((tmppath / "out/metadata.psv").exists())
+                with open(txt, encoding="utf8") as txt_f:
+                    non_blank_line_count = sum(1 for line in txt_f if line.strip())
+                for i in range(non_blank_line_count):
+                    self.assertTrue((tmppath / f"out/wavs/segment{i}.wav"))
+
+
+class MiscTests(TestCase):
+    def test_shell_complete(self):
+        self.assertEqual(complete_path(), [])
+        self.assertEqual(complete_path(None, None, None), [])
+
+    def test_segment(self):
+        segment = Segment("text", 500, 700, 0.42)
+        self.assertEqual(len(segment), 200)
+        self.assertEqual(repr(segment), "text (0.42): [ 500,  700)")

aligner/utils.py

@@ -32,7 +32,7 @@ def create_transducer(text, labels_dictionary, debug=False):
         if char not in allowable_chars and char not in fallback_mapping:
             fallback_mapping[char] = ""
     for k in fallback_mapping.keys():
-        if debug:
+        if debug:  # pragma: no cover
             print(
                 f"Found {k} which is not modelled by Wav2Vec2; skipping for alignment"
             )
@@ -49,7 +49,7 @@ def create_transducer(text, labels_dictionary, debug=False):
 
 
 def read_text(text_path):
-    with open(text_path) as f:
+    with open(text_path, encoding="utf8") as f:
         return [x.strip() for x in f]
 
 

requirements.txt

@@ -1,7 +1,9 @@
-torch>=2.1.0
-torchaudio>=2.1.0
 g2p>=1.0.20230417
+pydub>=0.23.1
 pympi-ling
-typer>=0.9.0
 rich>=10.11.0
 shellingham>=1.3.0
+soundfile>=0.10.2
+torch>=2.1.0
+torchaudio>=2.1.0
+typer>=0.9.0