Use external package for monotonic alignment search

.github/workflows/pypi-release.yml

@@ -7,8 +7,7 @@ defaults:
     shell:
       bash
 jobs:
-  build-sdist:
-    name: Build source distribution
+  build:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v4
@@ -20,37 +19,29 @@ jobs:
           if [[ "v$version" != "$tag" ]]; then
             exit 1
           fi
-      - uses: actions/setup-python@v5
+      - name: Install uv
+        uses: astral-sh/setup-uv@v3
         with:
-          python-version: 3.9
-      - run: |
-          python -m pip install -U pip setuptools build
-      - run: |
-          python -m build
-      - run: |
-          pip install dist/*.tar.gz
-      - uses: actions/upload-artifact@v4
-        with:
-          name: build-sdist
-          path: dist/*.tar.gz
-  build-wheels:
-    name: Build wheels on ${{ matrix.os }}
-    runs-on: ${{ matrix.os }}
-    strategy:
-      matrix:
-        os: [ubuntu-latest, windows-latest, macos-latest]
-    steps:
-      - uses: actions/checkout@v4
-      - name: Build wheels
-        uses: pypa/[email protected]
+          version: "0.4.27"
+          enable-cache: true
+          cache-dependency-glob: "**/pyproject.toml"
+      - name: Set up Python
+        run: uv python install 3.12
+      - name: Build sdist and wheel
+        run: uv build
+      - name: Test installation of sdist and wheel
+        run: |
+          uv venv --no-project
+          uv pip install dist/*.tar.gz
+          uv pip install dist/*.whl
       - uses: actions/upload-artifact@v4
         with:
-          name: build-wheels-${{ matrix.os }}
-          path: ./wheelhouse/*.whl
+          name: build
+          path: dist/*
   publish-artifacts:
     name: Publish to PyPI
     runs-on: ubuntu-latest
-    needs: [build-sdist, build-wheels]
+    needs: [build]
     environment:
       name: release
       url: https://pypi.org/p/coqui-tts
@@ -60,8 +51,7 @@ jobs:
       - uses: actions/download-artifact@v4
         with:
           path: dist
-          pattern: build-*
-          merge-multiple: true
+          pattern: build
       - run: |
           ls -lh dist/
       - name: Publish package distributions to PyPI

TTS/tts/layers/delightful_tts/acoustic_model.py

@@ -5,6 +5,7 @@
 import torch
 import torch.nn.functional as F
 from coqpit import Coqpit
+from monotonic_alignment_search import maximum_path
 from torch import nn
 
 from TTS.tts.layers.delightful_tts.conformer import Conformer
@@ -19,7 +20,7 @@
 from TTS.tts.layers.delightful_tts.pitch_adaptor import PitchAdaptor
 from TTS.tts.layers.delightful_tts.variance_predictor import VariancePredictor
 from TTS.tts.layers.generic.aligner import AlignmentNetwork
-from TTS.tts.utils.helpers import generate_path, maximum_path, sequence_mask
+from TTS.tts.utils.helpers import generate_path, sequence_mask
 
 logger = logging.getLogger(__name__)
 

TTS/tts/models/align_tts.py

@@ -3,6 +3,7 @@
 
 import torch
 from coqpit import Coqpit
+from monotonic_alignment_search import maximum_path
 from torch import nn
 from trainer.io import load_fsspec
 
@@ -12,7 +13,7 @@
 from TTS.tts.layers.feed_forward.encoder import Encoder
 from TTS.tts.layers.generic.pos_encoding import PositionalEncoding
 from TTS.tts.models.base_tts import BaseTTS
-from TTS.tts.utils.helpers import generate_path, maximum_path, sequence_mask
+from TTS.tts.utils.helpers import generate_path, sequence_mask
 from TTS.tts.utils.speakers import SpeakerManager
 from TTS.tts.utils.text.tokenizer import TTSTokenizer
 from TTS.tts.utils.visual import plot_alignment, plot_spectrogram

TTS/tts/models/forward_tts.py

@@ -4,6 +4,7 @@
 
 import torch
 from coqpit import Coqpit
+from monotonic_alignment_search import maximum_path
 from torch import nn
 from torch.cuda.amp.autocast_mode import autocast
 from trainer.io import load_fsspec
@@ -14,7 +15,7 @@
 from TTS.tts.layers.generic.pos_encoding import PositionalEncoding
 from TTS.tts.layers.glow_tts.duration_predictor import DurationPredictor
 from TTS.tts.models.base_tts import BaseTTS
-from TTS.tts.utils.helpers import average_over_durations, generate_path, maximum_path, sequence_mask
+from TTS.tts.utils.helpers import average_over_durations, generate_path, sequence_mask
 from TTS.tts.utils.speakers import SpeakerManager
 from TTS.tts.utils.text.tokenizer import TTSTokenizer
 from TTS.tts.utils.visual import plot_alignment, plot_avg_energy, plot_avg_pitch, plot_spectrogram

TTS/tts/models/glow_tts.py

@@ -4,6 +4,7 @@
 
 import torch
 from coqpit import Coqpit
+from monotonic_alignment_search import maximum_path
 from torch import nn
 from torch.cuda.amp.autocast_mode import autocast
 from torch.nn import functional as F
@@ -13,7 +14,7 @@
 from TTS.tts.layers.glow_tts.decoder import Decoder
 from TTS.tts.layers.glow_tts.encoder import Encoder
 from TTS.tts.models.base_tts import BaseTTS
-from TTS.tts.utils.helpers import generate_path, maximum_path, sequence_mask
+from TTS.tts.utils.helpers import generate_path, sequence_mask
 from TTS.tts.utils.speakers import SpeakerManager
 from TTS.tts.utils.synthesis import synthesis
 from TTS.tts.utils.text.tokenizer import TTSTokenizer

TTS/tts/models/vits.py

@@ -11,6 +11,7 @@
 import torchaudio
 from coqpit import Coqpit
 from librosa.filters import mel as librosa_mel_fn
+from monotonic_alignment_search import maximum_path
 from torch import nn
 from torch.cuda.amp.autocast_mode import autocast
 from torch.nn import functional as F
@@ -28,7 +29,7 @@
 from TTS.tts.layers.vits.stochastic_duration_predictor import StochasticDurationPredictor
 from TTS.tts.models.base_tts import BaseTTS
 from TTS.tts.utils.fairseq import rehash_fairseq_vits_checkpoint
-from TTS.tts.utils.helpers import generate_path, maximum_path, rand_segments, segment, sequence_mask
+from TTS.tts.utils.helpers import generate_path, rand_segments, segment, sequence_mask
 from TTS.tts.utils.languages import LanguageManager
 from TTS.tts.utils.speakers import SpeakerManager
 from TTS.tts.utils.synthesis import synthesis

TTS/tts/utils/helpers.py

@@ -3,13 +3,6 @@
 from scipy.stats import betabinom
 from torch.nn import functional as F
 
-try:
-    from TTS.tts.utils.monotonic_align.core import maximum_path_c
-
-    CYTHON = True
-except ModuleNotFoundError:
-    CYTHON = False
-
 
 class StandardScaler:
     """StandardScaler for mean-scale normalization with the given mean and scale values."""
@@ -168,73 +161,6 @@ def generate_path(duration, mask):
     return path
 
 
-def maximum_path(value, mask):
-    if CYTHON:
-        return maximum_path_cython(value, mask)
-    return maximum_path_numpy(value, mask)
-
-
-def maximum_path_cython(value, mask):
-    """Cython optimised version.
-    Shapes:
-        - value: :math:`[B, T_en, T_de]`
-        - mask: :math:`[B, T_en, T_de]`
-    """
-    value = value * mask
-    device = value.device
-    dtype = value.dtype
-    value = value.data.cpu().numpy().astype(np.float32)
-    path = np.zeros_like(value).astype(np.int32)
-    mask = mask.data.cpu().numpy()
-
-    t_x_max = mask.sum(1)[:, 0].astype(np.int32)
-    t_y_max = mask.sum(2)[:, 0].astype(np.int32)
-    maximum_path_c(path, value, t_x_max, t_y_max)
-    return torch.from_numpy(path).to(device=device, dtype=dtype)
-
-
-def maximum_path_numpy(value, mask, max_neg_val=None):
-    """
-    Monotonic alignment search algorithm
-    Numpy-friendly version. It's about 4 times faster than torch version.
-    value: [b, t_x, t_y]
-    mask: [b, t_x, t_y]
-    """
-    if max_neg_val is None:
-        max_neg_val = -np.inf  # Patch for Sphinx complaint
-    value = value * mask
-
-    device = value.device
-    dtype = value.dtype
-    value = value.cpu().detach().numpy()
-    mask = mask.cpu().detach().numpy().astype(bool)
-
-    b, t_x, t_y = value.shape
-    direction = np.zeros(value.shape, dtype=np.int64)
-    v = np.zeros((b, t_x), dtype=np.float32)
-    x_range = np.arange(t_x, dtype=np.float32).reshape(1, -1)
-    for j in range(t_y):
-        v0 = np.pad(v, [[0, 0], [1, 0]], mode="constant", constant_values=max_neg_val)[:, :-1]
-        v1 = v
-        max_mask = v1 >= v0
-        v_max = np.where(max_mask, v1, v0)
-        direction[:, :, j] = max_mask
-
-        index_mask = x_range <= j
-        v = np.where(index_mask, v_max + value[:, :, j], max_neg_val)
-    direction = np.where(mask, direction, 1)
-
-    path = np.zeros(value.shape, dtype=np.float32)
-    index = mask[:, :, 0].sum(1).astype(np.int64) - 1
-    index_range = np.arange(b)
-    for j in reversed(range(t_y)):
-        path[index_range, index, j] = 1
-        index = index + direction[index_range, index, j] - 1
-    path = path * mask.astype(np.float32)
-    path = torch.from_numpy(path).to(device=device, dtype=dtype)
-    return path
-
-
 def beta_binomial_prior_distribution(phoneme_count, mel_count, scaling_factor=1.0):
     P, M = phoneme_count, mel_count
     x = np.arange(0, P)