Monotonic Alignment Search

0.2.1 verified Mon Apr 27 auth: no python

Independent package implementing the monotonic alignment search algorithm from Glow-TTS for aligning text and speech. Currently at v0.2.1, supports PyTorch (CPU/GPU) and NumPy backends. Requires Python >=3.10.

pip install monotonic-alignment-search

Common errors

error ModuleNotFoundError: No module named 'monotonic_alignment_search' ↓

cause Package not installed or installed under a different name.

fix

pip install monotonic-alignment-search

error ImportError: cannot import name 'maximum_path_numpy' from 'monotonic_alignment_search' ↓

cause Attempting to use NumPy backend on v0.2.0+ where it was potentially removed or renamed.

fix

Use maximum_path (PyTorch backend) instead. If you need NumPy, downgrade to v0.1.1: pip install 'monotonic-alignment-search==0.1.1'

Warnings

gotcha Input attention must be non-negative and monotonic; the algorithm does not enforce this. Incorrect inputs may produce silent failures or suboptimal alignments. ↓

fix Ensure your attention matrix is monotonic and non-negative before calling maximum_path.

deprecated NumPy backend (maximum_path_numpy) was experimental in v0.1.1 and may be removed in a future release. ↓

fix Prefer the PyTorch backend (maximum_path) for stability. If you need NumPy, pin to v0.1.1.

breaking PyTorch backend removed or renamed? Not yet, but the v0.2.0 allowed choosing CPU/GPU; ensure your PyTorch installation is compatible. ↓

fix Install PyTorch separately according to your hardware (CPU/CUDA) before using this package.

Imports

maximum_path

wrong

from mas import maximum_path

correct

from monotonic_alignment_search import maximum_path

Old import from Glow-TTS; 'mas' is not a package here.

maximum_path_numpy
```
from monotonic_alignment_search import maximum_path_numpy
```
Only available since v0.1.1; use for NumPy backend.

Quickstart

Perform monotonic alignment search on a random hard attention matrix using PyTorch.

import torch
from monotonic_alignment_search import maximum_path

# Example: random hard attention (batch=1, 1 source, 5 target, no mask)
attn_hard = torch.rand(1, 1, 1, 5).cumsum(3).round().long()
# Perform MAS
path = maximum_path(attn_hard, attn_hard.new_zeros(attn_hard.shape))
print(path.shape)  # torch.Size([1, 1, 1, 5])