Add SIGMOSFilterStage for audio quality assessment by shubhamNvidia · Pull Request #1577 · NVIDIA-NeMo/Curator

shubhamNvidia · 2026-03-05T14:47:20Z

Adds SIGMOSFilterStage which filters audio based on SIGMOS (Signal-based Mean Opinion Score) quality metrics using an ONNX model. Predicts 7 quality dimensions (NOISE, OVRL, SIG, COL, DISC, LOUD, REVERB) on a 0-5 scale. Each threshold is independently configurable and can be set to None to disable. Items failing any active threshold are filtered out. Includes SIGMOSConfig with get_active_thresholds(), the sigmos_filter_module pipeline, and unit tests with mocked prediction function.

copy-pr-bot · 2026-03-05T14:47:25Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

greptile-apps · 2026-03-05T14:50:04Z

Greptile Summary

This PR adds SIGMOSFilterStage, an ONNX-based audio quality filter that scores seven P.804 MOS dimensions (NOISE, OVRL, SIG, COL, DISC, LOUD, REVERB) and drops segments below per-dimension thresholds. The model is auto-downloaded from GitHub on first use, cached at ~/.cache/nemo_curator/sigmos_model/, and GPU execution via ONNX Runtime's CUDAExecutionProvider is supported with a CPU fallback.

Confidence Score: 4/5

Safe to merge after addressing the non-atomic model download, which can leave a corrupt cache file that fails on subsequent runs without re-downloading.

Two P2 findings remain: the non-atomic/non-streaming download (where a partial write passes the size guard and causes an opaque ORT error) and the undeclared requests dependency in audio extras. These are worth fixing before merge, keeping the score at 4.

nemo_curator/stages/audio/filtering/sigmos.py (download logic); pyproject.toml (missing requests in audio_common)

Important Files Changed

Filename	Overview
nemo_curator/stages/audio/filtering/sigmos.py	Core filter stage with model download/setup logic; non-atomic download is a latent reliability issue
nemo_curator/stages/audio/filtering/sigmos_filter_module/third_party/sigmos/sigmos.py	Third-party SigMOS ONNX wrapper updated to use loguru; CUDAExecutionProvider fallback detection logic is correct
nemo_curator/utils/gpu_utils.py	Retained get_gpu_count/get_max_model_len_from_config; new ensure_cudnn_loaded uses glob to handle any libcudnn.so version
tests/stages/audio/filtering/test_sigmos.py	Unit tests cover pass/fail thresholds, None-disabled checks, nested segment mode, and missing audio inputs
pyproject.toml	requests used in sigmos.py but not declared in audio_common extras

Sequence Diagram

sequenceDiagram
    participant C as Caller
    participant S as SIGMOSFilterStage
    participant D as _download_model
    participant B as build_sigmos_model
    participant O as OnnxRuntime

    C->>S: setup(worker_metadata)
    S->>S: ensure_cudnn_loaded()
    S->>S: _initialize_model()
    S->>D: _resolve_model_path()
    D->>D: Check cache / download from GitHub
    D-->>S: model_path
    S->>B: build_sigmos_model(force_cpu, device_id, model_path)
    B->>O: InferenceSession(CUDA or CPU)
    O-->>B: session
    B-->>S: SigMOS instance
    S-->>C: ready

    C->>S: process(AudioTask)
    S->>S: _get_audio_numpy_sr()
    S->>O: model.run(audio, sr)
    O-->>S: MOS scores dict
    S->>S: _check_thresholds(scores)
    alt passes all thresholds
        S-->>C: AudioTask with sigmos_* fields
    else fails any threshold
        S-->>C: []
    end

_{Reviews (39): Last reviewed commit: "Adding Microsoft license to LICENSE file" | Re-trigger Greptile}

greptile-apps · 2026-03-05T14:50:10Z

+        _cudnn_loaded = True
+    except OSError:
+        logger.warning("Failed to load %s", cudnn_so, exc_info=True)
+        return False
+


Hardcoded cuDNN version libcudnn.so.9 will fail on other cuDNN versions

The function hardcodes libcudnn.so.9, which is specific to cuDNN 9.x. This will fail silently (returning False) on systems running cuDNN 8.x (e.g., libcudnn.so.8) or any future cuDNN 10+ version, even when cuDNN is correctly installed.

Consider searching for any available libcudnn.so.* in the directory:

import glob cudnn_candidates = sorted(glob.glob(os.path.join(cudnn_lib_dir, "libcudnn.so.*")), reverse=True) if not cudnn_candidates: logger.warning("No libcudnn.so.* found in %s", cudnn_lib_dir) return False cudnn_so = cudnn_candidates[0] # use newest available

greptile-apps · 2026-03-05T14:50:11Z

+# Model cache for storing initialized models by GPU ID
+_MODEL_CACHE = {}


Global _MODEL_CACHE is not thread-safe

The module-level _MODEL_CACHE dictionary is shared across all threads. In a multi-worker environment (e.g., Ray actors on the same process, or Python threading), concurrent SIGMOSPipeline initializations can race on the if cache_key not in _MODEL_CACHE check and build the same model multiple times, or read a partially-constructed model.

Consider protecting it with a lock:

import threading _MODEL_CACHE = {} _MODEL_CACHE_LOCK = threading.Lock()

And wrapping the check-and-insert:

with _MODEL_CACHE_LOCK: if cache_key not in _MODEL_CACHE: _MODEL_CACHE[cache_key] = build_sigmos_model(...) self.model = _MODEL_CACHE[cache_key]

sarahyurick · 2026-03-10T17:20:33Z

+
+
+@dataclass
+class SIGMOSConfig:


Discussed offline. We do not need a config class per stage.

sarahyurick · 2026-03-10T17:22:33Z

+        options.inter_op_num_threads = 1
+        options.intra_op_num_threads = 1
+
+        use_gpu = True if is_gpu_support_available() and not force_cpu else False


This can be handled by the stage instead.

sarahyurick · 2026-03-10T17:37:29Z

+    def test_defaults(self) -> None:
+        cfg = SIGMOSConfig()
+        assert cfg.noise_threshold == 4.0
+        assert cfg.ovrl_threshold == 3.5
+        assert cfg.sig_threshold is None
+        assert cfg.col_threshold is None
+        assert cfg.disc_threshold is None
+        assert cfg.loud_threshold is None
+        assert cfg.reverb_threshold is None
+
+    def test_from_dict(self) -> None:
+        cfg = SIGMOSConfig.from_dict({"noise_threshold": 3.0, "sig_threshold": 2.5})
+        assert cfg.noise_threshold == 3.0
+        assert cfg.sig_threshold == 2.5
+        assert cfg.ovrl_threshold == 3.5
+
+    def test_from_dict_none(self) -> None:
+        cfg = SIGMOSConfig.from_dict(None)
+        assert cfg.noise_threshold == 4.0
+
+    def test_from_dict_ignores_unknown(self) -> None:
+        cfg = SIGMOSConfig.from_dict({"unknown": 99, "noise_threshold": 1.0})
+        assert cfg.noise_threshold == 1.0


We can remove these.

sarahyurick · 2026-03-10T17:37:46Z

+        )
+        assert cfg.get_active_thresholds() == {}
+
+    def test_get(self) -> None:


This can be removed.

sarahyurick · 2026-03-10T17:37:55Z

+    def test_stage_properties(self) -> None:
+        stage = SIGMOSFilterStage()
+        assert stage.name == "SIGMOSFilter"
+        assert stage.inputs() == (["data"], [])
+        _, output_keys = stage.outputs()
+        for key in ["sigmos_noise", "sigmos_ovrl", "sigmos_sig", "sigmos_col",
+                     "sigmos_disc", "sigmos_loud", "sigmos_reverb"]:
+            assert key in output_keys
+
+    def test_config_overrides_params(self) -> None:
+        cfg = SIGMOSConfig(noise_threshold=2.0, ovrl_threshold=None, sig_threshold=3.0)
+        stage = SIGMOSFilterStage(config=cfg)
+        assert stage.noise_threshold == 2.0
+        assert stage.ovrl_threshold is None
+        assert stage.sig_threshold == 3.0


We can remove these.

sarahyurick · 2026-03-10T17:38:22Z

+        assert len(result.data) == 0
+
+    @patch("nemo_curator.stages.audio.filtering.sigmos.SIGMOSFilterStage._initialize_model")
+    def test_preserves_task_metadata(self, mock_init) -> None:


We can remove this.

…_audio_mos, simplified pipeline

…th logger

…ction, remove shebang and __main__, rewrite SIGMOS tests to match stage API

sarahyurick

Kicking off the tests and Ruff formatter.

sarahyurick · 2026-03-23T19:24:21Z

+# ---------------------------------------------------------------------------
+
+TEST_DATA_DIR = Path(
+    "/lustre/fsw/portfolios/maxine/users/shbhawsar/debug/nemo_curator"


We can't have this in the codebase.

This code will not be the part of pr its for internal testing, will remove this

sarahyurick · 2026-03-23T19:29:12Z

/ok to test 3d7a714

shubhamNvidia · 2026-04-01T20:53:29Z

/ok to test 672dfe0

…lection time

shubhamNvidia · 2026-04-01T21:24:45Z

/ok to test a1f3c82

shubhamNvidia · 2026-04-02T11:32:58Z

/ok to test 3747312

shubhamNvidia · 2026-04-02T16:49:45Z

/ok to test e089988

sarahyurick · 2026-04-02T18:23:26Z

Following up from our call, this can be deleted.

sarahyurick · 2026-04-02T18:23:39Z

Following up from our call, this can be deleted.

sarahyurick · 2026-04-02T18:25:05Z

+        Use .with_(resources=Resources(gpus=X)) to configure GPU allocation.
+    """
+
+    model_path: str = "model/model-sigmos_1697718653_41d092e8-epo-200.onnx"


Suggested change

model_path: str = "model/model-sigmos_1697718653_41d092e8-epo-200.onnx"

model_path: str

sarahyurick · 2026-04-02T18:25:58Z

+            "sigmos_reverb",
+        ]
+
+    def setup(self, _: WorkerMetadata | None = None) -> None:


The setup_on_node function can verify whether model_path exists, and if not, raise an error along with a message about where it can be downloaded from.

Done. setup_on_node now auto-downloads the ONNX model from Microsoft's SIG-Challenge repo into ~/.cache/nemo_curator/sigmos_model/. If download fails it raises RuntimeError with the source URL and expected local path so the user can download manually. Users can also skip auto-download entirely by passing model_path= directly.

shubhamNvidia · 2026-04-02T20:45:13Z

/ok to test b3228af

Signed-off-by: shbhawsar <shbhawsar@nvidia.com>

ayushdg · 2026-04-03T14:11:27Z

@@ -0,0 +1,121 @@
+import logging


License header

ayushdg · 2026-04-03T14:11:39Z

@@ -0,0 +1,33 @@
+# What is SIGMOS?


License header

Done please review

ayushdg · 2026-04-03T14:17:52Z

+import onnxruntime as ort
+import scipy
+
+logger = logging.getLogger(__name__)


Our repo uses loguru's logger throughout for logging, maybe that's a modification that can be made here.

ayushdg · 2026-04-03T14:22:54Z

+    current_dir = os.path.dirname(os.path.abspath(__file__))
+    sys.path.append(os.path.join(current_dir, "third_party"))
+    from sigmos.sigmos import build_sigmos_model


I wouldn't recommend sys path hacks to account for imports. If import errors are common we can fix the top level import to be more stable

ayushdg · 2026-04-03T14:24:24Z

+        return audio, int(sample_rate)
+
+    path = item.get("audio_filepath")
+    if path and os.path.isfile(path):


not blocking for this release since it's a pattern across audio but we should make this more cloud friendly.

ayushdg · 2026-04-03T14:28:43Z

+            return None
+
+        try:
+            score_data = self._predict_audio_mos(audio_np, sample_rate, model_path=self._resolve_model_path())


Each predict call will instantiate a new class object and load the cached model? Wouldn't it be better to ensure the model is loaded once per actor and predict is called throughout?

ayushdg · 2026-04-03T14:29:36Z

+        logger.debug(
+            "nvidia-cudnn-cu12 is not installed; "
+            "cuDNN must be available on the system LD_LIBRARY_PATH for GPU inference."
+        )


Should this be a warning or a debug call? (I'm not sure asking for curiosity)

Kept as debug because the pip package is only one of several ways to provide cuDNN — system installs are common. The actual GPU fallback is already surfaced as a warning in the ONNX session setup.

ayushdg · 2026-04-03T14:30:16Z

+    ld_path = os.environ.get("LD_LIBRARY_PATH", "")
+    if cudnn_lib_dir not in ld_path:
+        os.environ["LD_LIBRARY_PATH"] = cudnn_lib_dir + (":" + ld_path if ld_path else "")
+
+    # Eagerly load cuDNN shared libraries into the process address space.
+    # Setting LD_LIBRARY_PATH alone is not enough once the process has started
+    # because the dynamic linker caches its search paths at startup.
+    # ONNX Runtime's CUDA provider uses dlopen() for sub-libraries like
+    # libcudnn_adv.so.9, so we must pre-load all of them.
+    import glob
+
+    cudnn_libs = sorted(glob.glob(os.path.join(cudnn_lib_dir, "libcudnn*.so*")))
+    if not cudnn_libs:
+        logger.warning("No libcudnn*.so* files found in %s", cudnn_lib_dir)
+        return False
+
+    # Load the main library first (other libs depend on it), then the rest.
+    # The main library matches "libcudnn.so.<version>" (no underscore after "libcudnn").
+    cudnn_libs.sort(key=lambda p: (not os.path.basename(p).startswith("libcudnn.so."), p))
+
+    for lib_path in cudnn_libs:
+        try:
+            ctypes.cdll.LoadLibrary(lib_path)
+            logger.debug("Pre-loaded %s", lib_path)
+        except OSError:  # noqa: PERF203
+            logger.warning("Failed to load %s", lib_path, exc_info=True)
+
+    _cudnn_loaded = True
+
+    return True


This sort of logic can be fragile in a lot of deployment setups, we should open an issue to track if we can improve this behavior.

ayushdg · 2026-04-03T14:37:07Z

+    )
+
+
+class TestSIGMOSFilterStage:


Is it possible to have a test that uses the real model?

If not will the benchmark handle both CPU and GPU fallbacks for this stage?

The ONNX model will be removed from the repo (users download from source), so a real-model unit test would need download infrastructure in CI, adding fragility and latency to every test run. Unit tests use mocks to cover all stage logic (thresholds, nested segments, error paths); real-model end-to-end validation will be added to the benchmark pipeline, same pattern as FLEURS/ASR. The ONNX layer already handles CPU/GPU fallback transparently tries CUDAExecutionProvider first, falls back to CPUExecutionProvider if unavailable.

- Merge upstream/main, resolve __init__.py conflicts keeping SIGMOS + UTMOS + SpeakerSep - Add MIT (Microsoft) + NVIDIA license headers to third-party sigmos files - Switch third-party sigmos.py from logging to loguru for consistency - Remove sys.path hack in sigmos_pipeline.py, use direct relative import - Refactor SIGMOSFilterStage: load model once in setup(), call self._model.run() directly - Update tests to mock _initialize_model and use _make_mock_model helper - Fix Ruff lint (__all__ sorting) Signed-off-by: shbhawsar <shbhawsar@nvidia.com>

shubhamNvidia · 2026-04-05T09:19:27Z

/ok to test 75809cc

- Remove ONNX model files from repo (third_party/sigmos/ and model/) - Add auto-download from Microsoft's SIG-Challenge GitHub repository - Model cached at ~/.cache/nemo_curator/sigmos_model/ by default - Users can override with model_path= or model_dir= - Add setup_on_node() for multi-node pre-download - Add file validation (missing/empty check with clear error message) - Uses requests.get() matching NSFW filter pattern Signed-off-by: shbhawsar <shbhawsar@nvidia.com>

mohammadaaftabv · 2026-04-05T11:49:10Z

git history commit 5faad7a (added), 75809cc (removed). Models no longer in latest HEAD of PR, but 50 MB still in git history.
Follow-up from merged PR #1577 (SIGMOSFilterStage):

Two identical copies of the SIGMOS ONNX model were checked into git:

sigmos_filter_module/model/model-sigmos_1697718653_41d092e8-epo-200.onnx (25 MB)
sigmos_filter_module/third_party/sigmos/model-sigmos_1697718653_41d092e8-epo-200.onnx (25 MB)

That's 50 MB of binary permanently in git history. Every git clone of NeMo Curator downloads this even if users never use SIGMOS.

Follow-up needed:

Remove at least one duplicate immediately
Ideally host the model externally (NGC or HuggingFace) and download on first use — similar to how UTMOSFilterStage loads its model via torch.hub

mohammadaaftabv · 2026-04-05T11:50:12Z

+import librosa
+import numpy as np
+import onnxruntime as ort
+import scipy


The third-party sigmos.py (Microsoft SigMOS) does import librosa and import scipy at module level (lines 7-8 of sigmos_filter_module/third_party/sigmos/sigmos.py). Neither librosa nor scipy is declared in pyproject.toml — not in base deps, not in any audio extras.

Impact:

from nemo_curator.stages.audio.filtering import SIGMOSFilterStage crashes with ModuleNotFoundError: No module named 'librosa'

Since filtering/__init__.py imports SIGMOSFilterStage at module level, the entire nemo_curator.stages.audio.filtering package is broken without librosa installed

This also breaks AudioDataFilterStage (this PR) since it imports from filtering

Fix: add librosa and scipy to the audio_cpu and/or audio_cuda12 extras in pyproject.toml.
This I think will come up when running unit tests in this pr ci/cd as well.

mohammadaaftabv · 2026-04-05T11:51:24Z

+
+    def teardown(self) -> None:
+        self._model = None
+        torch.cuda.empty_cache()


SIGMOSFilterStage.teardown() in nemo_curator/stages/audio/filtering/sigmos.py calls torch.cuda.empty_cache() unconditionally:

def teardown(self) -> None: self._model = None torch.cuda.empty_cache() # crashes if CUDA not available

The unconditional torch.cuda.empty_cache() in teardown() now affects 4 of 6 new stages:

Stage PR teardown() guarded?

SIGMOSFilterStage #1577 ❌ No

VADSegmentationStage #1578 ❌ No

UTMOSFilterStage #1639 ❌ No

SpeakerSeparationStage #1579 ❌ No

BandFilterStage #1576 ✅ Yes

MonoConversionStage #1575 N/A (no GPU)

BandFilterStage (#1576) already has the correct pattern:

if torch.cuda.is_available(): torch.cuda.empty_cache()

mohammadaaftabv · 2026-04-05T11:52:51Z

+
+from .third_party.sigmos.sigmos import build_sigmos_model
+
+_MODEL_CACHE: dict[str, Any] = {}


_SIGMOSPipeline and predict_audio_mos() with a module-level _MODEL_CACHE dict, but SIGMOSFilterStage never imports or uses any of it. The stage has its own _initialize_model() that directly calls build_sigmos_model().

This file seems like dead code. Remove it, or if model caching across workers was intended, wire it into the stage.
Please let me know if this file is used somewhere. Also add a readme edit somewhere relevant so that it is clear to users what exactly is being done by each added stage/file.

Thanks for bringing this up. Removed this file

mohammadaaftabv · 2026-04-05T12:10:28Z

+
+    def teardown(self) -> None:
+        self._model = None
+        torch.cuda.empty_cache()


The unconditional torch.cuda.empty_cache() in teardown() now affects 4 of 6 new stages:

Stage PR teardown() guarded?

SIGMOSFilterStage #1577 ❌ No

VADSegmentationStage #1578 ❌ No

UTMOSFilterStage #1639 ❌ No

SpeakerSeparationStage #1579 ❌ No

BandFilterStage #1576 ✅ Yes

MonoConversionStage #1575 N/A (no GPU)

BandFilterStage (#1576) already has the correct pattern:

if torch.cuda.is_available(): torch.cuda.empty_cache()

shubhamNvidia · 2026-04-05T13:58:15Z

/ok to test d0936c4

- Guard torch.cuda.empty_cache() with torch.cuda.is_available() in SIGMOSFilterStage.teardown() to prevent crash on CPU-only machines - Add librosa and scipy to audio_common extras in pyproject.toml (required by third-party sigmos.py module-level imports) - Remove unused sigmos_pipeline.py (dead code, never imported by stage) - Regenerate uv.lock

Signed-off-by: Sarah Yurick <53962159+sarahyurick@users.noreply.github.com>

sarahyurick · 2026-04-06T03:06:11Z

/ok to test f297f01

shubhamNvidia · 2026-04-06T16:25:12Z

/ok to test 9aa21a3

…clause

Signed-off-by: shubhamNvidia <shbhawsar@nvidia.com>

shubhamNvidia · 2026-04-06T17:00:11Z

/ok to test 961e8d0

shubhamNvidia added 3 commits March 5, 2026 02:24

Add SIGMOSFilterStage and ensure_cudnn_loaded utility

2c80669

Add SIGMOS model files

5faad7a

Add unit tests for SIGMOSFilterStage

e64f75b

github-actions Bot added the community-request label Mar 5, 2026

greptile-apps Bot reviewed Mar 5, 2026

View reviewed changes

sarahyurick self-requested a review March 5, 2026 18:10

ayushdg removed the community-request label Mar 5, 2026

sarahyurick requested changes Mar 10, 2026

View reviewed changes

shubhamNvidia added 2 commits March 11, 2026 03:45

Simplified config, removed temp file creation, uses in-memory predict…

f3f8928

…_audio_mos, simplified pipeline

Guard GPU fallback, delegate GPU detection to stage, replace print wi…

01dc37e

…th logger

github-actions Bot added the community-request label Mar 12, 2026

shubhamNvidia force-pushed the pr/audio-sigmos branch 2 times, most recently from c02ad9b to b76b6ad Compare March 20, 2026 06:53

Fix copyrights, restore upstream gpu_utils and added ensure cudnn fun…

ffd5888

…ction, remove shebang and __main__, rewrite SIGMOS tests to match stage API

shubhamNvidia force-pushed the pr/audio-sigmos branch from b76b6ad to ffd5888 Compare March 20, 2026 07:52

shubhamNvidia added 2 commits March 22, 2026 08:39

Add common utilities, replace duplicate _resolve_model_path in sigmos.py

be96268

change resource allocation pattern

ab83471

sarahyurick reviewed Mar 23, 2026

View reviewed changes

copy-pr-bot Bot temporarily deployed to test March 23, 2026 19:29 Inactive

copy-pr-bot Bot had a problem deploying to nemo-ci March 23, 2026 19:29 Failure

copy-pr-bot Bot temporarily deployed to nemo-ci March 23, 2026 19:29 Inactive

copy-pr-bot Bot had a problem deploying to nemo-ci March 23, 2026 19:29 Failure

copy-pr-bot Bot temporarily deployed to nemo-ci March 23, 2026 19:29 Inactive

fix: lazy import predict_audio_mos to avoid onnxruntime import at col…

672dfe0

…lection time

shubhamNvidia added 2 commits April 1, 2026 21:36

fix: guard sigmos import to avoid ImportError during test collection

a1f3c82

Merge branch 'main' into pr/audio-sigmos

3747312

Merge remote-tracking branch 'upstream/main' into pr/audio-sigmos

e089988

sarahyurick reviewed Apr 2, 2026

View reviewed changes

Merge branch 'main' into pr/audio-sigmos

b3228af

Merge upstream/main, resolve conflicts keeping both changes

32215b4

Signed-off-by: shbhawsar <shbhawsar@nvidia.com>

ayushdg previously requested changes Apr 3, 2026

View reviewed changes

sarahyurick mentioned this pull request Apr 3, 2026

Add ADV audio processing dependencies #1574

Merged

mohammadaaftabv requested changes Apr 5, 2026

View reviewed changes

mohammadaaftabv previously requested changes Apr 5, 2026

View reviewed changes

shubhamNvidia and others added 2 commits April 5, 2026 14:05

Merge branch 'main' into pr/audio-sigmos

f297f01

Signed-off-by: Sarah Yurick <53962159+sarahyurick@users.noreply.github.com>

shubhamNvidia added 3 commits April 6, 2026 16:33

Add Microsoft MIT attribution for SIGMOS and Apache 2.0 contribution …

b194a3d

…clause

Merge branch 'main' into pr/audio-sigmos

9aa21a3

Adding Microsoft license to LICENSE file

961e8d0

Signed-off-by: shubhamNvidia <shbhawsar@nvidia.com>

sarahyurick approved these changes Apr 6, 2026

View reviewed changes

		# Model cache for storing initialized models by GPU ID
		_MODEL_CACHE = {}

	model_path: str = "model/model-sigmos_1697718653_41d092e8-epo-200.onnx"
	model_path: str

Stage	PR	`teardown()` guarded?
SIGMOSFilterStage	#1577	❌ No
VADSegmentationStage	#1578	❌ No
UTMOSFilterStage	#1639	❌ No
SpeakerSeparationStage	#1579	❌ No
BandFilterStage	#1576	✅ Yes
MonoConversionStage	#1575	N/A (no GPU)


		from .third_party.sigmos.sigmos import build_sigmos_model

		_MODEL_CACHE: dict[str, Any] = {}

Uh oh!

Conversation

shubhamNvidia commented Mar 5, 2026

Uh oh!

copy-pr-bot Bot commented Mar 5, 2026

Uh oh!

greptile-apps Bot commented Mar 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Greptile Summary

Confidence Score: 4/5

Important Files Changed

Sequence Diagram

Uh oh!

Uh oh!

greptile-apps Bot Mar 5, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps Bot Mar 5, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

sarahyurick left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

sarahyurick commented Mar 23, 2026

Uh oh!

shubhamNvidia commented Apr 1, 2026

Uh oh!

shubhamNvidia commented Apr 1, 2026

Uh oh!

shubhamNvidia commented Apr 2, 2026

Uh oh!

shubhamNvidia commented Apr 2, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

shubhamNvidia commented Apr 2, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

greptile-apps Bot commented Mar 5, 2026 •

edited

Loading