Skip to main content

Fixed an issue where the inference progress hook reported a higher number of completions than the total

https://github.com/pyannote/pyannote-audio/pull/2074

Purpose​

Sliding inference (Inference) calls the progress hook as completed=c + self.batch_size each time a batch is processed. Even if the last batch is smaller than batch_size, since batch_size is added as-is, the hook receives a value indicating that “more than the total has been completed” unless the number of chunks is a multiple of batch_size. Subsequently, if the last chunk reports the correct value, the progress may even reverse.

These are the (completed, total) values received by the hook in the develop branch (duration=1.0, step=0.5).

| Audio Length | batch_size | Call Value |

| --- | --- | --- |

| 3.0 seconds | 32 | (0, 5), (32, 5) |

| 3.05 seconds | 32 | (0, 6), (32, 6), (6, 6) |

| 3.05 seconds | 4 | (0, 6), (4, 6), (8, 6), (6, 6) |

| 10.0 seconds | 4 | (0, 19), (4, 19), …, (16, 19), (20, 19) |

With the default value of batch_size=32, files shorter than 32 chunks are processed starting from the first batch. Since the progress of the segmentation stage in the speaker segmentation pipeline is tracked by this hook, hooks that calculate ratios or determine completion using completed >= total will receive incorrect values.

Changes​

  • src/pyannote/audio/core/inference.py — Restricted completed to min(c + self.batch_size, num_chunks) (the actual number of chunks processed)
  • tests/test_inference_hook.py — Added a new test using a small dummy model to verify four scenarios. Checks that the progress bar does not exceed 100%, does not move backward, and starts at 0 and ends at 100%.
  • CHANGELOG.md — Added one line to the next entry

Tests​

$ pytest tests/test_inference_hook.py     # Before modification: 4 failed
$ pytest tests/test_inference_hook.py     # After modification: 4 passed

These are tests run without a pre-trained model. The local environment is Python 3.12, torch 2.8.0 (CPU). The remaining tests that require a model could not be run locally.

Duplicate Check​

I searched for open and closed issues and PRs using the progress hook, completed, and total, and found no duplicate issues. Open PRs (#2048, #1992) that touch inference.py do not change this line.

Results​

Submitted to the develop branch on October 6, 2026. Awaiting review.