[aviv1ron1]

closes #165

Problem

Some ASR models emit a marker such as <|nospeech|> for silence or non-speech. That marker was spliced into the LLM prompt as if the user had said it.

Change

• New config field asr_no_speech_markers: list[str] | None, set at compose time with a repeatable --asr-no-speech-marker flag. The marker depends on the ASR model, so there is no default: nothing is removed unless configured. Empty markers and a bare string (instead of a list) are rejected.
• Every occurrence of each configured marker is removed from each transcribed window, before long-audio windows are merged. This covers a marker next to speech, repeated markers, a marker inside one window, and the self_chunks=True path.
• Only the configured exact strings are removed. Any other tag (timings, emotions, language tags) is kept.
• The markers are part of the transcriber cache key; the processor passes them through from config.
• Documented in docs/AUDIO.md (compose flag and config.json form).

Supersedes #169, which only dropped a transcript that was exactly <|nospeech|> and hardcoded that string.

Usage

python -m granite_switch.composer.compose_granite_switch \
  --adapters ... --enable-audio \
  --asr-no-speech-marker "<|nospeech|>"

Existing checkpoints: add "asr_no_speech_markers": ["<|nospeech|>"] to config.json.

Tests

• New: marker removal (alone, repeated, next to speech, mid-sentence, several markers, other tags kept, no markers configured, self_chunks=True, per-window before merge), cache key, config round-trip and validation, CLI flag parsing, processor passthrough.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[aviv1ron1]

/gpu-test-audio

[github-actions]

✅ GPU tests passed — vllm30-audio

221 passed, 1 skipped, 2908 deselected, 23 warnings in 668.23s (0:11:08)

Commit: c390236c211c0889fdee2bdbed9ac05a97428714
Full run & artifact log

Last 40 log lines

tests/composer/test_compose_e2e.py:83
  /tmp/granite-switch/tests/composer/test_compose_e2e.py:83: PytestUnknownMarkWarning: Unknown pytest.mark.xdist_group - is this a typo?  You can register custom marks to avoid this warning - for details, see https://docs.pytest.org/en/stable/how-to/mark.html
    @pytest.mark.xdist_group("compose_e2e")

tests/composer/test_multi_audio_compose_e2e.py:124
  /tmp/granite-switch/tests/composer/test_multi_audio_compose_e2e.py:124: PytestUnknownMarkWarning: Unknown pytest.mark.xdist_group - is this a typo?  You can register custom marks to avoid this warning - for details, see https://docs.pytest.org/en/stable/how-to/mark.html
    @pytest.mark.xdist_group("multi_audio_compose_e2e")

tests/composer/test_multi_audio_compose_e2e.py:130
  /tmp/granite-switch/tests/composer/test_multi_audio_compose_e2e.py:130: PytestUnknownMarkWarning: Unknown pytest.mark.xdist_group - is this a typo?  You can register custom marks to avoid this warning - for details, see https://docs.pytest.org/en/stable/how-to/mark.html
    @pytest.mark.xdist_group("multi_audio_compose_e2e")

tests/composer/test_multi_audio_compose_e2e.py:230
  /tmp/granite-switch/tests/composer/test_multi_audio_compose_e2e.py:230: PytestUnknownMarkWarning: Unknown pytest.mark.xdist_group - is this a typo?  You can register custom marks to avoid this warning - for details, see https://docs.pytest.org/en/stable/how-to/mark.html
    @pytest.mark.xdist_group("multi_audio_compose_e2e")

tests/composer/test_upstream_files.py:218
  /tmp/granite-switch/tests/composer/test_upstream_files.py:218: PytestUnknownMarkWarning: Unknown pytest.mark.xdist_group - is this a typo?  You can register custom marks to avoid this warning - for details, see https://docs.pytest.org/en/stable/how-to/mark.html
    @pytest.mark.xdist_group("upstream_build_e2e")

.venv/lib/python3.12/site-packages/torch/jit/_script.py:365: 14 warnings
  /tmp/granite-switch/.venv/lib/python3.12/site-packages/torch/jit/_script.py:365: DeprecationWarning: `torch.jit.script_method` is deprecated. Please switch to `torch.compile` or `torch.export`.
    warnings.warn(

tests/unit/test_asr.py::TestMonoAndResample::test_resample_real_vllm_backend[8000-16000]
tests/unit/test_asr.py::TestMonoAndResample::test_resample_real_vllm_backend[44100-16000]
  /tmp/granite-switch/.venv/lib/python3.12/site-packages/torch/backends/mkldnn/__init__.py:96: UserWarning: TF32 acceleration on top of oneDNN is available for Intel GPUs. The current Torch version does not have Intel GPU Support. (Triggered internally at /__w/pytorch/pytorch/aten/src/ATen/Context.cpp:265.)
    torch._C._set_onednn_allow_tf32(_allow_tf32)

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
=== 221 passed, 1 skipped, 2908 deselected, 23 warnings in 668.23s (0:11:08) ===
sys:1: DeprecationWarning: builtin type swigvarlink has no __module__ attribute
===== ALL GPU TESTS PASSED =====
[14:11:00] <job> Succeeded
[14:11:02] verified: found success sentinel in pod log
[14:11:02] cleanup: deleting <job> (exit=0)
<job> "<job>" deleted

Verdict: PASSED

[github-actions]

✅ GPU tests passed — vllm26-audio

222 passed, 2908 deselected, 21 warnings in 1084.13s (0:18:04)

Commit: c390236c211c0889fdee2bdbed9ac05a97428714
Full run & artifact log

Last 40 log lines

<frozen importlib._bootstrap>:488
  <frozen importlib._bootstrap>:488: DeprecationWarning: builtin type SwigPyPacked has no __module__ attribute

<frozen importlib._bootstrap>:488
  <frozen importlib._bootstrap>:488: DeprecationWarning: builtin type SwigPyObject has no __module__ attribute

tests/composer/test_compose_e2e.py:83
  /tmp/granite-switch/tests/composer/test_compose_e2e.py:83: PytestUnknownMarkWarning: Unknown pytest.mark.xdist_group - is this a typo?  You can register custom marks to avoid this warning - for details, see https://docs.pytest.org/en/stable/how-to/mark.html
    @pytest.mark.xdist_group("compose_e2e")

tests/composer/test_multi_audio_compose_e2e.py:124
  /tmp/granite-switch/tests/composer/test_multi_audio_compose_e2e.py:124: PytestUnknownMarkWarning: Unknown pytest.mark.xdist_group - is this a typo?  You can register custom marks to avoid this warning - for details, see https://docs.pytest.org/en/stable/how-to/mark.html
    @pytest.mark.xdist_group("multi_audio_compose_e2e")

tests/composer/test_multi_audio_compose_e2e.py:130
  /tmp/granite-switch/tests/composer/test_multi_audio_compose_e2e.py:130: PytestUnknownMarkWarning: Unknown pytest.mark.xdist_group - is this a typo?  You can register custom marks to avoid this warning - for details, see https://docs.pytest.org/en/stable/how-to/mark.html
    @pytest.mark.xdist_group("multi_audio_compose_e2e")

tests/composer/test_multi_audio_compose_e2e.py:230
  /tmp/granite-switch/tests/composer/test_multi_audio_compose_e2e.py:230: PytestUnknownMarkWarning: Unknown pytest.mark.xdist_group - is this a typo?  You can register custom marks to avoid this warning - for details, see https://docs.pytest.org/en/stable/how-to/mark.html
    @pytest.mark.xdist_group("multi_audio_compose_e2e")

tests/composer/test_upstream_files.py:218
  /tmp/granite-switch/tests/composer/test_upstream_files.py:218: PytestUnknownMarkWarning: Unknown pytest.mark.xdist_group - is this a typo?  You can register custom marks to avoid this warning - for details, see https://docs.pytest.org/en/stable/how-to/mark.html
    @pytest.mark.xdist_group("upstream_build_e2e")

.venv/lib/python3.12/site-packages/torch/jit/_script.py:365: 14 warnings
  /tmp/granite-switch/.venv/lib/python3.12/site-packages/torch/jit/_script.py:365: DeprecationWarning: `torch.jit.script_method` is deprecated. Please switch to `torch.compile` or `torch.export`.
    warnings.warn(

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
======== 222 passed, 2908 deselected, 21 warnings in 1084.13s (0:18:04) ========
sys:1: DeprecationWarning: builtin type swigvarlink has no __module__ attribute
===== ALL GPU TESTS PASSED =====
[14:19:00] <job> Succeeded
[14:19:03] verified: found success sentinel in pod log
[14:19:03] cleanup: deleting <job> (exit=0)
<job> "<job>" deleted

Verdict: PASSED