Whisper transcribe() integration test skipped due to a known, uninvestigated failure

[msluszniak]

Description

S2TTranscribeTests.TranscribeReturnsValidChars in common/rnexecutorch/tests/integration/SpeechToTextTest.cpp:84 is disabled with:

GTEST_SKIP() << "TODO: known failure on this branch; needs investigation.";

This is the only test that asserts the full SpeechToText::transcribe() path returns a well-formed result (language == "en", task == "transcribe", non-empty segments, non-empty text, ASCII chars). Its sibling tests — EncodeReturnsNonNull, EmptyResultOnSilence, InvalidLanguageThrows — are not skipped, and the skip reason is explicitly not attributed to the emulator (unlike the TextToImage skips). So a genuine defect in transcribe() is currently masked.

Secondary, minor: the char-range validation in the same test is half-dead —

EXPECT_GE(static_cast<unsigned char>(c), 0);  // always true: unsigned char is >= 0
EXPECT_LE(static_cast<unsigned char>(c), 127);

The >= 0 assertion is a tautology. Either drop it or check the raw signed char/byte if the intent is to catch non-ASCII output. Worth cleaning up when the test is re-enabled.

Steps to reproduce

1. Build and run the native integration test suite under common/rnexecutorch/tests/integration.
2. Remove the GTEST_SKIP() at SpeechToTextTest.cpp:84.
3. Observe S2TTranscribeTests.TranscribeReturnsValidChars fail.
4. Investigate which assertion fails (language / duration / task / segments / text) to localize the regression in transcribe().

Snack or a link to a repository

N/A — reproduced via the in-repo native test suite.

React Native Executorch version

main @ v0.8.4

React Native version

N/A — native (C++) integration test, not React-Native-version dependent.

Platforms

Platform agnostic (shared C++ in common/) — confirm whether the failure reproduces on both iOS and Android backends.

JavaScript runtime / Workflow / Architecture / Build type / Device

N/A — native gtest.

AI model

Whisper

Acknowledgements

Yes — searched existing issues; the open speech-to-text issues (#769 qwen3-asr, #553 speech-to-speech) are about new models, not this transcribe failure.