Transformers Community Adds Whisper Runtime Reproduction; Single-Element EOS List May Interrupt Timestamp Generation

The Transformers community raised an issue with Whisper timestamp generation on October 13 and added a reproduction using a real model the following day. The reported environment was Windows, Python 3.13.1, and Transformers 5.19.0.dev0. When generation is configured with both return_timestamps=True and eos_token_id=[50257], the timestamp processor may raise a TypeError, interrupting transcription. This report concerns a specific parameter combination; the scope of affected released versions has not been established. Issue and discussion

The problem occurs in score masking during generation: according to the report, WhisperTimeStampLogitsProcessor uses the EOS token ID directly as a slice boundary, and a Python list cannot be used as a slice index. The official documentation explains that logits processors adjust generation scores, while Whisper’s timestamp processor constrains the valid structure of timestamp output. The engineering checks, therefore, are how generation settings reach model-specific processors and whether each layer makes consistent assumptions about data types. Generation utilities documentation

The discussion also highlights open-source projects’ expectations for evidence supporting a fix. A maintainer asked whether the issue occurred in actual use and said static-analysis concerns submitted by coding agents had added review burden, asking for runtime evidence with a commonly used model first. The reporter then said they had reproduced the original code’s failure and the modified code’s success with openai/whisper-tiny and real audio, and that 49 tests in the relevant test file passed. These results remain the author’s account. Maintainer and author replies

The Whisper Tiny model card documents using return_timestamps=True to obtain segment-level timestamps, giving deployers a clear feature path to verify. Teams using shared wrappers for generation settings can compare integer EOS values with single-element lists in their own versions and check the timestamp output. The author said they had proposed a fix to normalize single-element inputs; follow-up should track maintainer confirmation, merge status, and the release version. The report does not establish that default transcription workflows are generally broken. Model card