[mikemikimike]

Fixes #165

Summary

• Treat the <|nospeech|> control token emitted by multilingual CTC checkpoints as an empty transcript.
• Keep the existing empty-transcript placeholder behavior so silent or non-speech audio cannot add the control token to the LLM prompt.
• Add a regression test covering surrounding whitespace and the normalized empty result.

Tests

• uv run --frozen --no-default-groups --extra hf --extra compose python -m pytest tests/unit/test_asr.py::TestTranscriber -q
• uvx pre-commit run --files src/granite_switch/vllm/audio/asr.py tests/unit/test_asr.py
• git diff --check

[mikemikimike]

The focused regression tests for this change pass. The remaining failure in tests/unit/test_asr.py::TestPatchedPipelineRestores::test_pipeline_attribute_is_restored was reproduced on the base commit 1cc5225, so it is pre-existing and is not introduced by this PR. The Windows environment also cannot collect the full unit suite because the project’s GPU-only vllm dependency has no supported Windows installation.