Abstract
Time-aligned speech annotation is typically done in Praat and ELAN, tools developed prior to the current Python-based NLP ecosystem. Taking advantage of moden NLP workflows means exporting files and reimporting the results. We address this with Besra, an open source desktop application for editing and analysing time-aligned annotation, written entirely in Python as a headless core with a cross-platform PySide6 interface. It is extensible, as Plugins are ordinary Python files executed in the host interpreter that receive the running application as an argument. The extension surface is every distribution installed in the environment, including the machine learning stack and the standards infrastructure of comparative linguistics. Waveform, spectrogram, pitch, formant, amplitude, and annotation displays share one time axis and one selection. It offers with a wider choice of estimation methods than annotation editors usually offer. Real-time quality measures are drawn on the same as the annotation is made, including a per-interval score for suitability as speech recognition training data, and a heatmap flagging transcriptions unlike the rest of their tier. Both are computed from the recording and its own annotation, so they need no lexicon or pretrained model and work on undescribed languages. Besra reads and writes Praat TextGrid, ELAN.eaf, and a documented version-controllable tab-separated format, with round-trip fidelity under test, so a project can adopt it for a single task. Besra is released under the GNU Affero General Public License version 3, developed on Codeberg and archived on Zenodo.