Merge pull request #147 from iautolab/feat/replace-deepfilternet-with-dpdfnet

feat: replace DeepFilterNet with DPDFNet for noise suppression

Author: MaleicAcid

.github/workflows/ci.yml

@@ -14,7 +14,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python-version: [ "3.10", "3.11", "3.12" ]
+        python-version: [ "3.11", "3.12" ]
         os: [ windows-latest, ubuntu-latest ]
 
     defaults:

.github/workflows/dispatch_CI.yml

@@ -14,7 +14,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python-version: [ "3.9" ]
+        python-version: [ "3.11" ]
         os: [ macos-latest ]
 
     defaults:

README.md

@@ -59,7 +59,7 @@ into `.lrc` subtitles with LLMs such as
     ```
 
 5. **(Optional)** If you need noise suppression (`noise_suppress=True`), install the full extras
-   which includes torch and DeepFilterNet:
+   which includes DPDFNet:
 
     ```shell
     pip install 'openlrc[full]'
@@ -76,7 +76,7 @@ into `.lrc` subtitles with LLMs such as
 OpenLRC keeps several package-root APIs lightweight to import.
 
 The following imports are guaranteed not to eagerly load heavyweight runtime dependencies such as
-`torch`, `spacy`, `faster-whisper`, `tiktoken`, or `lingua`:
+`dpdfnet`, `spacy`, `faster-whisper`, `tiktoken`, or `lingua`:
 
 ```python
 import openlrc
@@ -91,15 +91,16 @@ without immediately starting transcription or language-processing work.
 Heavy dependencies are loaded only when the corresponding features are first used. For example:
 
 - `faster-whisper` is loaded when transcription is first needed.
-- `torch` and `df.enhance` are loaded when noise suppression is used.
+- `dpdfnet` is loaded when noise suppression is used.
 - `spacy` is loaded when sentence segmentation or related NLP helpers are used.
 - `tiktoken` is loaded when token counting is used.
 - `lingua` is loaded when language detection helpers are used.
 
 > [!NOTE]
-> The base `pip install openlrc` does **not** include torch or DeepFilterNet.
-> These are only installed with `pip install 'openlrc[full]'` and are only needed
-> for noise suppression (`noise_suppress=True`).
+> The base `pip install openlrc` does **not** include dpdfnet.
+> It is only installed with `pip install 'openlrc[full]'` and is only needed
+> for noise suppression (`noise_suppress=True`). On first use, DPDFNet downloads
+> its model (default `dpdfnet2_48khz_hr`, ~10MB) to `~/.cache/dpdfnet/models`.
 
 ## Usage 🐍
 

openlrc/defaults.py

@@ -40,7 +40,9 @@
     "speech_pad_ms": 400,
 }
 
-default_preprocess_options = {"atten_lim_db": 15}
+DEFAULT_DPDFNET_MODEL = "dpdfnet2_48khz_hr"
+
+default_preprocess_options = {"atten_lim_db": 15, "dpdfnet_model": DEFAULT_DPDFNET_MODEL}
 
 # File name suffixes used throughout the pipeline.
 # The processing chain produces files like:

openlrc/media_utils.py

@@ -4,7 +4,7 @@
 """Media-related utility functions that depend on heavy external libraries.
 
 Functions in this module import packages such as ``ffmpeg``, ``filetype``,
-``audioread``, ``torch``, and ``spacy`` inside their bodies.  Keeping them
+``audioread``, and ``spacy`` inside their bodies.  Keeping them
 separate from :mod:`openlrc.utils` ensures that the lightweight translation
 path never triggers those imports — critical for Nuitka ``--nofollow-import-to``
 builds where the heavy packages are intentionally excluded.
@@ -14,7 +14,7 @@
 
 import subprocess
 from pathlib import Path
-from typing import TYPE_CHECKING, Any
+from typing import TYPE_CHECKING
 
 if TYPE_CHECKING:
     from spacy.language import Language as SpacyLanguage
@@ -84,16 +84,6 @@ def get_audio_duration(path: str | Path) -> float:
         return audio.duration
 
 
-def release_memory(model: Any) -> None:
-    try:
-        import torch
-    except ImportError:
-        return
-
-    if isinstance(model, torch.nn.Module):
-        torch.cuda.empty_cache()
-
-
 def get_spacy_lib(lang):
     special_case = {"core_web": ["zh", "en"], "ent_wiki": ["xx"]}
 

openlrc/models.py

@@ -1,20 +1,11 @@
 #  Copyright (C) 2025. Hao Zheng
 #  All rights reserved.
 
-import sys
 from dataclasses import dataclass
-from enum import Enum
+from enum import StrEnum
 
-if sys.version_info >= (3, 11):
-    from enum import StrEnum as _StrEnum
-else:
-    class _StrEnum(str, Enum):
-        """Backport of StrEnum for Python 3.10."""
-        def __str__(self) -> str:
-            return self.value
 
-
-class ModelProvider(_StrEnum):
+class ModelProvider(StrEnum):
     ANTHROPIC = "anthropic"
     OPENAI = "openai"
     GOOGLE = "google"

openlrc/preprocess.py

@@ -7,9 +7,14 @@
 from ffmpeg_normalize import FFmpegNormalize
 from tqdm import tqdm
 
-from openlrc.defaults import LOUDNORM_SUFFIX, NOISE_SUPPRESSED_SUFFIX, PREPROCESSED_DIR, default_preprocess_options
+from openlrc.defaults import (
+    DEFAULT_DPDFNET_MODEL,
+    LOUDNORM_SUFFIX,
+    NOISE_SUPPRESSED_SUFFIX,
+    PREPROCESSED_DIR,
+    default_preprocess_options,
+)
 from openlrc.logger import logger
-from openlrc.media_utils import release_memory
 from openlrc.utils import get_preprocessed_path
 
 
@@ -63,17 +68,15 @@ def noise_suppression(self, audio_paths: list[Path], atten_lim_db: int = 15):
             return []
 
         try:
-            import torch
-            from df.enhance import enhance, init_df, load_audio, save_audio
+            import dpdfnet
+            import librosa
+            import soundfile
         except ImportError:
-            raise ImportError(
-                "Noise suppression requires torch and deepfilternet. Install them with: pip install 'openlrc[full]'"
-            )
+            raise ImportError("Noise suppression requires dpdfnet. Install it with: pip install 'openlrc[full]'")
 
         if "atten_lim_db" in self.options:
             atten_lim_db = self.options["atten_lim_db"]
-
-        model, df_state, _ = init_df()
+        model = self.options.get("dpdfnet_model", DEFAULT_DPDFNET_MODEL)
         chunk_size = 180  # 3 min
 
         ns_audio_paths = []
@@ -82,31 +85,33 @@ def noise_suppression(self, audio_paths: list[Path], atten_lim_db: int = 15):
             ns_path = output_path / f"{audio_name}{NOISE_SUPPRESSED_SUFFIX}.wav"
 
             if not ns_path.exists():
-                audio, info = load_audio(str(audio_path), sr=df_state.sr())
+                waveform, sr = librosa.load(str(audio_path), sr=None, mono=False)
+                sr = int(sr)
 
                 # Split audio into 3 min chunks
+                chunk_samples = chunk_size * sr
                 audio_chunks = [
-                    audio[:, i : i + chunk_size * info.sample_rate]
-                    for i in range(0, audio.shape[1], chunk_size * info.sample_rate)
+                    waveform[..., i : i + chunk_samples] for i in range(0, waveform.shape[-1], chunk_samples)
                 ]
 
-                enhanced_chunks = []
-                for ac in tqdm(audio_chunks, desc=f"Noise suppressing for {audio_name}"):
-                    enhanced_chunks.append(enhance(model, df_state, ac, atten_lim_db=atten_lim_db))
-
-                enhanced = torch.cat(enhanced_chunks, dim=1)
+                try:
+                    with soundfile.SoundFile(str(ns_path), mode="w", samplerate=sr, channels=1, subtype="PCM_16") as f:
+                        for ac in tqdm(audio_chunks, desc=f"Noise suppressing for {audio_name}"):
+                            enhanced = dpdfnet.enhance(ac, sr, model=model, attn_limit_db=atten_lim_db)
 
-                if enhanced.shape != audio.shape:
-                    raise ValueError(
-                        f"Enhanced audio shape does not match original audio shape: {enhanced.shape} != {audio.shape}"
-                    )
+                            if enhanced.shape[-1] != ac.shape[-1]:
+                                raise ValueError(
+                                    f"Enhanced audio shape does not match original audio shape: "
+                                    f"{enhanced.shape} != {ac.shape}"
+                                )
 
-                save_audio(str(ns_path), enhanced, sr=df_state.sr())
+                            f.write(enhanced)
+                except Exception:
+                    ns_path.unlink(missing_ok=True)
+                    raise
 
             ns_audio_paths.append(ns_path)
 
-        release_memory(model)
-
         return ns_audio_paths
 
     def loudness_normalization(self, audio_paths: list[Path]):

pyproject.toml

@@ -7,7 +7,7 @@ name = "openlrc"
 version = "1.7.0a1"
 description = "Transcribe (whisper) and translate (gpt) voice into LRC file."
 authors = [{ name = "Hao Zheng", email = "zhenghaosustc@gmail.com" }]
-requires-python = ">=3.10, <3.13"
+requires-python = ">=3.11, <3.13"
 readme = "README.md"
 license = "MIT"
 keywords = ["openai-gpt3", "whisper", "voice transcribe", "lrc"]
@@ -38,19 +38,17 @@ dependencies = [
     "ffmpeg-normalize>=1.27.5,<2",
     "google-genai>=1.70.0",
     "json_repair==0.25.2",
-    "onnxruntime>=1.20.0,<1.24; python_version < '3.11'",
-    "onnxruntime>=1.20.0,<2; python_version >= '3.11'",
+    "onnxruntime>=1.20.0,<2",
     "pip>=25.1",
 ]
 
 [project.optional-dependencies]
-# Noise suppression via DeepFilterNet (requires torch).
+# Noise suppression via DPDFNet (pure Python + ONNX Runtime, no torch/torchaudio).
 # Install with: pip install openlrc[full]
 # Only needed when using noise_suppress=True in LRCer.run() or Preprocessor.run().
 full = [
-    "torch>=2.6.0",
-    "torchaudio>=2.0.0",
-    "deepfilternet>=0.5.6,<0.6",
+    "dpdfnet>=0.6.0",
+    "numba>=0.58.0",  # first version supporting both Python 3.11 and 3.12
 ]
 litellm = [
     "litellm>=1.60,<1.85",
@@ -75,21 +73,8 @@ name = "PyPI"
 url = "https://pypi.org/simple/"
 default = true
 
-[[tool.uv.index]]
-name = "pytorch-cu124"
-url = "https://download.pytorch.org/whl/cu124"
-explicit = true
-
-[tool.uv.sources]
-torch = [
-    { index = "pytorch-cu124", marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
-]
-torchaudio = [
-    { index = "pytorch-cu124", marker = "sys_platform == 'linux' or sys_platform == 'win32'" },
-]
-
 [tool.ruff]
-target-version = "py310"
+target-version = "py311"
 line-length = 120
 
 [tool.ruff.format]
@@ -106,7 +91,7 @@ known-first-party = ["openlrc"]
 split-on-trailing-comma = false
 
 [tool.pyright]
-pythonVersion = "3.10"
+pythonVersion = "3.11"
 exclude = ["openlrc/gui_streamlit"]
 
 [tool.hatch.build.targets.sdist]

tests/test_lazy_imports.py

@@ -10,7 +10,7 @@
 
 
 class TestLazyImports(unittest.TestCase):
-    FORBIDDEN_ROOTS = {"faster_whisper", "spacy", "tiktoken", "torch", "lingua"}
+    FORBIDDEN_ROOTS = {"faster_whisper", "spacy", "tiktoken", "dpdfnet", "lingua"}
     PROBE_MARKER = "__OPENLRC_LAZY_IMPORTS__="
 
     def _loaded_modules_after(self, statement: str):
@@ -59,7 +59,7 @@ def test_translate_path_does_not_load_media_utils_or_heavy_deps(self):
         # The translate path legitimately loads lingua (via validators.py) and
         # tiktoken (via utils.get_text_token_number).  Only check for the truly
         # heavy packages that would bloat a Nuitka binary.
-        heavy_roots = {"faster_whisper", "spacy", "torch"}
+        heavy_roots = {"faster_whisper", "spacy", "dpdfnet"}
         loaded = self._loaded_modules_after(
             "from openlrc.agents import create_chatbot; "
             "from openlrc.context import TranslateInfo; "