Revert "Revert "Use speex for noise suppression and auto gain"" (#124637)

Revert "Revert "Use speex for noise suppression and auto gain" (#124620)"

This reverts commit 302ffe5e56488f2f139686d375e5c82339304c87.

homeassistant/components/assist_pipeline/audio_enhancer.py

@@ -5,6 +5,7 @@ from dataclasses import dataclass
 import logging
 
 from pymicro_vad import MicroVad
+from pyspeex_noise import AudioProcessor
 
 from .const import BYTES_PER_CHUNK
 
@@ -41,8 +42,8 @@ class AudioEnhancer(ABC):
         """Enhance chunk of PCM audio @ 16Khz with 16-bit mono samples."""
 
 
-class MicroVadEnhancer(AudioEnhancer):
+class MicroVadSpeexEnhancer(AudioEnhancer):
-    """Audio enhancer that just runs microVAD."""
+    """Audio enhancer that runs microVAD and speex."""
 
     def __init__(
         self, auto_gain: int, noise_suppression: int, is_vad_enabled: bool
@@ -50,6 +51,24 @@ class MicroVadEnhancer(AudioEnhancer):
         """Initialize audio enhancer."""
         super().__init__(auto_gain, noise_suppression, is_vad_enabled)
 
+        self.audio_processor: AudioProcessor | None = None
+
+        # Scale from 0-4
+        self.noise_suppression = noise_suppression * -15
+
+        # Scale from 0-31
+        self.auto_gain = auto_gain * 300
+
+        if (self.auto_gain != 0) or (self.noise_suppression != 0):
+            self.audio_processor = AudioProcessor(
+                self.auto_gain, self.noise_suppression
+            )
+            _LOGGER.debug(
+                "Initialized speex with auto_gain=%s, noise_suppression=%s",
+                self.auto_gain,
+                self.noise_suppression,
+            )
+
         self.vad: MicroVad | None = None
         self.threshold = 0.5
 
@@ -61,12 +80,17 @@ class MicroVadEnhancer(AudioEnhancer):
         """Enhance 10ms chunk of PCM audio @ 16Khz with 16-bit mono samples."""
         is_speech: bool | None = None
 
+        assert len(audio) == BYTES_PER_CHUNK
+
         if self.vad is not None:
             # Run VAD
-            assert len(audio) == BYTES_PER_CHUNK
             speech_prob = self.vad.Process10ms(audio)
             is_speech = speech_prob > self.threshold
 
+        if self.audio_processor is not None:
+            # Run noise suppression and auto gain
+            audio = self.audio_processor.Process10ms(audio).audio
+
         return EnhancedAudioChunk(
             audio=audio, timestamp_ms=timestamp_ms, is_speech=is_speech
         )

homeassistant/components/assist_pipeline/manifest.json

@@ -7,5 +7,5 @@
   "integration_type": "system",
   "iot_class": "local_push",
   "quality_scale": "internal",
-  "requirements": ["pymicro-vad==1.0.1"]
+  "requirements": ["pymicro-vad==1.0.1", "pyspeex-noise==1.0.0"]
 }

homeassistant/components/assist_pipeline/pipeline.py

@@ -49,7 +49,7 @@ from homeassistant.util import (
 )
 from homeassistant.util.limited_size_dict import LimitedSizeDict
 
-from .audio_enhancer import AudioEnhancer, EnhancedAudioChunk, MicroVadEnhancer
+from .audio_enhancer import AudioEnhancer, EnhancedAudioChunk, MicroVadSpeexEnhancer
 from .const import (
     BYTES_PER_CHUNK,
     CONF_DEBUG_RECORDING_DIR,
@@ -589,7 +589,7 @@ class PipelineRun:
         # Initialize with audio settings
         if self.audio_settings.needs_processor and (self.audio_enhancer is None):
             # Default audio enhancer
-            self.audio_enhancer = MicroVadEnhancer(
+            self.audio_enhancer = MicroVadSpeexEnhancer(
                 self.audio_settings.auto_gain_dbfs,
                 self.audio_settings.noise_suppression_level,
                 self.audio_settings.is_vad_enabled,

homeassistant/components/voip/voip.py

@@ -33,7 +33,7 @@ from homeassistant.components.assist_pipeline import (
 )
 from homeassistant.components.assist_pipeline.audio_enhancer import (
     AudioEnhancer,
-    MicroVadEnhancer,
+    MicroVadSpeexEnhancer,
 )
 from homeassistant.components.assist_pipeline.vad import (
     AudioBuffer,
@@ -235,7 +235,7 @@ class PipelineRtpDatagramProtocol(RtpDatagramProtocol):
         try:
             # Wait for speech before starting pipeline
             segmenter = VoiceCommandSegmenter(silence_seconds=self.silence_seconds)
-            audio_enhancer = MicroVadEnhancer(0, 0, True)
+            audio_enhancer = MicroVadSpeexEnhancer(0, 0, True)
             chunk_buffer: deque[bytes] = deque(
                 maxlen=self.buffered_chunks_before_speech,
             )

homeassistant/package_constraints.txt

@@ -49,6 +49,7 @@ pymicro-vad==1.0.1
 PyNaCl==1.5.0
 pyOpenSSL==24.2.1
 pyserial==3.5
+pyspeex-noise==1.0.0
 python-slugify==8.0.4
 PyTurboJPEG==1.7.1
 pyudev==0.24.1

requirements_all.txt

@@ -2228,6 +2228,9 @@ pysoma==0.0.12
 # homeassistant.components.spc
 pyspcwebgw==0.7.0
 
+# homeassistant.components.assist_pipeline
+pyspeex-noise==1.0.0
+
 # homeassistant.components.squeezebox
 pysqueezebox==0.7.1
 

requirements_test_all.txt

@@ -1782,6 +1782,9 @@ pysoma==0.0.12
 # homeassistant.components.spc
 pyspcwebgw==0.7.0
 
+# homeassistant.components.assist_pipeline
+pyspeex-noise==1.0.0
+
 # homeassistant.components.squeezebox
 pysqueezebox==0.7.1