Release 3.4.0

Bug Fixes

• Blank references inside a list or numpy array are now rejected by error_handler(), matching the existing check for a single blank reference string. The message names the index of the first blank reference, and the wer(), wers(), werp(), werps(), summary() and summaryp() functions print that message and return None. Previously a blank reference inside a sequence contributed a Word Error Rate of 0.0.

• Added a zero guard to the per-row weighted Word Error Rate calculation in summaryp(), matching the guard already used in werps(), so a reference with zero words no longer raises an uncaught ZeroDivisionError.

Enhancements

• Corrected the docstrings of wer(), wers(), werp(), werps(), summary() and summaryp() to state that the functions print a message and return None when the input is invalid. The return type annotation of wer() is now float | None. In version 4.0.0 invalid input will raise an exception instead of returning None.

• Pinned the build requirements in pyproject.toml to exact versions of meson-python, meson and Cython, and removed wheel from the build requirements.

• The compiled extension is now built against a stated NumPy C API version (NPY_TARGET_VERSION=NPY_1_25_API_VERSION), the API level provided by NumPy 1.26.

• Added a linetrace Meson build option, off by default, that compiles metrics.pyx with Cython line tracing for coverage measurement.

• The default Meson build type is now release with b_ndebug=if-release, and the license in meson.build is declared with the SPDX identifier BSD-3-Clause.

• Added the <cnp.int32_t> casts to the two buffer assignments in _calculations_wer_only_reuse_ptr() that did not carry them, matching the other calculation paths, and replaced the metrics.pyx module docstring with a description of the three calculation paths and their return shapes.

• The Windows wheel no longer contains the MSVC import library (.lib) of the compiled extension.

• The source distribution no longer contains the CI configuration, documentation, benchmark and development files.

• The generated werpy/metrics.c file produced by a line tracing build is ignored.

• Every token of a reference and hypothesis pair is now mapped to one canonical object per distinct word before the edit distance is computed, so the dynamic programming loops in metrics.pyx compare words by pointer equality instead of by string comparison. The results of every calculation path are unchanged.

• When the reference and the hypothesis are the same string, the calculation functions return a zero edit distance and zero counts directly, without building the token mapping or the edit distance table.

• The Word Error Rate batch path no longer splits every hypothesis in a preliminary pass to size its buffers. The row buffers are sized on demand as the batch is processed.

• The Word Error Rate batch path now shares one token buffer across every pair in the batch, grown on demand and freed once when the batch is finished, instead of allocating and freeing a buffer for each pair.

• The punctuation translation table used by normalize() is now built once when the module is imported rather than on every call.

• Added support for Python 3.14. The package builds and passes its test suite on CPython 3.14, wheels for CPython 3.14 are now built for Linux, macOS and Windows, and the Programming Language :: Python :: 3.14 classifier was added to pyproject.toml.

• Added werpy/metrics.pyi, a stub file for the compiled werpy.metrics module, installed with the package alongside py.typed. It declares the parameter and return types of calculations(), metrics(), calculations_fast(), metrics_fast(), calculations_wer_only() and metrics_wer_only().

• Added type annotations to normalize(), error_handler(), wer(), wers(), werp(), werps(), summary() and summaryp(). The reference and hypothesis parameters are annotated as str, list[str] or a NumPy array of strings, the types accepted by error_handler(). wers() and werps() declare a float return for a pair of strings and a list[float] return for a pair of lists or arrays, and normalize() declares a str return for a string and a list[str] return for a list, tuple or array.

• Linux wheels are now built on the manylinux_2_28 image and require glibc 2.28 or newer. Previously they were built on the manylinux2014 image, which requires glibc 2.17.