Noise suppression tweaks.

README.md

@@ -44,7 +44,7 @@ Barnard includes real-time noise suppression for microphone input to filter out
 
 ### Features
 - **Real-time processing**: Noise suppression is applied during audio capture with minimal latency
-- **Configurable threshold**: Adjustable noise gate threshold (default: 0.02)
+- **Configurable amount**: Adjustable suppression amount via threshold value (default: `0.08`)
 - **Persistent settings**: Noise suppression preferences are saved in your configuration file
 - **Multiple control methods**: Toggle via hotkey, command line flag, or FIFO commands
 
@@ -57,10 +57,11 @@ Barnard includes real-time noise suppression for microphone input to filter out
 ### Configuration Example
 ```toml
 noisesuppressionenabled = true
-noisesuppressionthreshold = 0.02
+noisesuppressionthreshold = 0.08
 ```
 
-The noise suppression algorithm uses a combination of high-pass filtering and noise gating to reduce unwanted background sounds while preserving voice quality.
+`noisesuppressionthreshold` accepts values from `0.0` to `1.0`, where higher values apply stronger suppression.
+The noise suppression algorithm uses adaptive noise-floor tracking, transient suppression, and smoothed gain reduction to reduce background noise while preserving voice quality.
 
 ## FIFO Control
 

config/user_config.go

@@ -128,7 +128,7 @@ func (c *Config) LoadConfig() {
 		jc.NoiseSuppressionEnabled = &enabled
 	}
 	if c.config.NoiseSuppressionThreshold == nil {
-		threshold := float32(0.02)
+		threshold := float32(0.08)
 		jc.NoiseSuppressionThreshold = &threshold
 	}
 	if c.config.VoiceEffect == nil {
@@ -249,12 +249,25 @@ func (c *Config) SetNoiseSuppressionEnabled(enabled bool) {
 
 func (c *Config) GetNoiseSuppressionThreshold() float32 {
 	if c.config.NoiseSuppressionThreshold == nil {
-		return 0.02
+		return 0.08
 	}
-	return *c.config.NoiseSuppressionThreshold
+	threshold := *c.config.NoiseSuppressionThreshold
+	if threshold < 0.0 {
+		return 0.0
+	}
+	if threshold > 1.0 {
+		return 1.0
+	}
+	return threshold
 }
 
 func (c *Config) SetNoiseSuppressionThreshold(threshold float32) {
+	if threshold < 0.0 {
+		threshold = 0.0
+	}
+	if threshold > 1.0 {
+		threshold = 1.0
+	}
 	c.config.NoiseSuppressionThreshold = &threshold
 	c.SaveConfig()
 }

noise/suppression.go

@@ -2,6 +2,7 @@ package noise
 
 import (
 	"math"
+	"sync"
 )
 
 // Ensure Suppressor implements the NoiseProcessor interface
@@ -12,129 +13,200 @@ var _ interface {
 
 // Suppressor handles noise suppression for audio samples
 type Suppressor struct {
-	enabled    bool
-	threshold  float32
-	gainFactor float32
-	
-	// Simple high-pass filter state for DC removal
+	mu sync.Mutex
+
+	enabled   bool
+	threshold float32
+
+	// High-pass filter state for low-frequency rumble/DC removal.
 	prevInput  float32
 	prevOutput float32
-	alpha      float32
-	
-	// Click detection state
-	clickThreshold    float32
-	clickDecay        float32
-	recentClickEnergy float32
+	hpAlpha    float32
+
+	// Adaptive suppression state.
+	envelope        float32
+	noiseFloor      float32
+	suppressionGain float32
+	clickEnergy     float32
+
+	// Tunables.
+	envelopeAttack  float32
+	envelopeRelease float32
+	noiseAttack     float32
+	noiseRelease    float32
+	gainAttack      float32
+	gainRelease     float32
+	speechRatio     float32
+	clickDecay      float32
+	minNoiseFloor   float32
 }
 
 // NewSuppressor creates a new noise suppressor
 func NewSuppressor() *Suppressor {
-	return &Suppressor{
-		enabled:           false,
-		threshold:         0.01,     // Reduced noise threshold level for less aggressive filtering
-		gainFactor:        0.9,      // Less aggressive gain reduction for noise
-		alpha:             0.98,     // More stable high-pass filter coefficient
-		clickThreshold:    0.15,     // Threshold for detecting keyboard clicks
-		clickDecay:        0.95,     // How quickly click energy decays
-		recentClickEnergy: 0.0,      // Tracks recent click activity
+	s := &Suppressor{
+		enabled:         false,
+		threshold:       0.08,
+		hpAlpha:         0.995,
+		envelopeAttack:  0.18,
+		envelopeRelease: 0.02,
+		noiseAttack:     0.08,
+		noiseRelease:    0.002,
+		gainAttack:      0.35,
+		gainRelease:     0.02,
+		speechRatio:     4.0,
+		clickDecay:      0.93,
+		minNoiseFloor:   0.0008,
+		suppressionGain: 1.0,
 	}
+	s.resetStateLocked()
+	return s
 }
 
 // SetEnabled enables or disables noise suppression
 func (s *Suppressor) SetEnabled(enabled bool) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if s.enabled == enabled {
+		return
+	}
 	s.enabled = enabled
+	s.resetStateLocked()
 }
 
 // IsEnabled returns whether noise suppression is enabled
 func (s *Suppressor) IsEnabled() bool {
+	s.mu.Lock()
+	defer s.mu.Unlock()
 	return s.enabled
 }
 
 // SetThreshold sets the noise threshold (0.0 to 1.0)
 func (s *Suppressor) SetThreshold(threshold float32) {
-	if threshold >= 0.0 && threshold <= 1.0 {
-		s.threshold = threshold
-	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.threshold = clampFloat32(threshold, 0.0, 1.0)
 }
 
 // GetThreshold returns the current noise threshold
 func (s *Suppressor) GetThreshold() float32 {
+	s.mu.Lock()
+	defer s.mu.Unlock()
 	return s.threshold
 }
 
 // ProcessSamples applies noise suppression to audio samples
 func (s *Suppressor) ProcessSamples(samples []int16) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
 	if !s.enabled || len(samples) == 0 {
 		return
 	}
 
-	// Calculate frame energy for click detection
-	var frameEnergy float32 = 0.0
-	for _, sample := range samples {
-		floatSample := float32(sample) / 32767.0
-		frameEnergy += floatSample * floatSample
-	}
-	frameEnergy = float32(math.Sqrt(float64(frameEnergy / float32(len(samples)))))
-	
-	// Detect sudden energy spikes (likely keyboard clicks)
-	energySpike := frameEnergy - s.recentClickEnergy
-	isClick := energySpike > s.clickThreshold && frameEnergy > 0.05
-	
-	// Update recent click energy with decay
-	s.recentClickEnergy = s.recentClickEnergy*s.clickDecay + frameEnergy*(1.0-s.clickDecay)
-	
-	// Improved noise suppression algorithm
+	intensity := s.thresholdToIntensity()
+	minGain := 1.0 - (0.92 * intensity)
+	eps := float32(1e-6)
+
 	for i, sample := range samples {
-		// Convert to float for processing
-		floatSample := float32(sample) / 32767.0
-		
-		// Apply high-pass filter for DC removal
-		filtered := s.highPassFilter(floatSample)
-		
-		// Calculate signal strength (RMS-like)
-		strength := float32(math.Abs(float64(filtered)))
-		
-		// Apply noise gate with smooth transition
-		var gainReduction float32 = 1.0
-		
-		// If we detected a click, apply stronger suppression
-		if isClick {
-			gainReduction = s.gainFactor * 0.3 // Much stronger reduction for clicks
-		} else if strength < s.threshold {
-			// Normal noise gate for low-level sounds
-			gainReduction = strength / s.threshold
-			if gainReduction < s.gainFactor {
-				gainReduction = s.gainFactor
-			}
-		}
-		
-		// Apply gain reduction
-		processed := filtered * gainReduction
-		
-		// Convert back to int16 with proper clipping
-		processedInt := processed * 32767.0
-		if processedInt > 32767 {
-			processedInt = 32767
-		} else if processedInt < -32767 {
-			processedInt = -32767
-		}
-		
-		samples[i] = int16(processedInt)
+		floatSample := float32(sample) / 32768.0
+		filtered := s.highPassFilterLocked(floatSample)
+		absSample := float32(math.Abs(float64(filtered)))
+
+		s.updateEnvelopeLocked(absSample)
+		s.updateNoiseFloorLocked()
+
+		snr := s.envelope / (s.noiseFloor + eps)
+		voicePresence := clampFloat32((snr-1.0)/(s.speechRatio-1.0), 0.0, 1.0)
+
+		targetGain := minGain + ((1.0 - minGain) * voicePresence)
+		targetGain = s.applyTransientSuppressionLocked(absSample, voicePresence, minGain, targetGain)
+
+		s.applyGainSmoothingLocked(targetGain)
+
+		processed := filtered * s.suppressionGain
+		processed = clampFloat32(processed, -1.0, 1.0)
+		samples[i] = int16(processed * 32767.0)
 	}
 }
 
-// highPassFilter applies a simple high-pass filter to remove DC component
-func (s *Suppressor) highPassFilter(input float32) float32 {
+func (s *Suppressor) highPassFilterLocked(input float32) float32 {
 	// Simple high-pass filter: y[n] = alpha * (y[n-1] + x[n] - x[n-1])
-	output := s.alpha * (s.prevOutput + input - s.prevInput)
+	output := s.hpAlpha * (s.prevOutput + input - s.prevInput)
 	s.prevInput = input
 	s.prevOutput = output
 	return output
 }
 
+func (s *Suppressor) thresholdToIntensity() float32 {
+	// Keep lower legacy threshold values meaningful while allowing up to very aggressive suppression.
+	return 1.0 - float32(math.Exp(float64(-28.0*clampFloat32(s.threshold, 0.0, 1.0))))
+}
+
+func (s *Suppressor) updateEnvelopeLocked(absSample float32) {
+	if absSample > s.envelope {
+		s.envelope += s.envelopeAttack * (absSample - s.envelope)
+	} else {
+		s.envelope += s.envelopeRelease * (absSample - s.envelope)
+	}
+	if s.envelope < s.minNoiseFloor {
+		s.envelope = s.minNoiseFloor
+	}
+}
+
+func (s *Suppressor) updateNoiseFloorLocked() {
+	coef := s.noiseRelease
+	if s.envelope < s.noiseFloor*2.2 {
+		coef = s.noiseAttack
+	}
+	s.noiseFloor += coef * (s.envelope - s.noiseFloor)
+	if s.noiseFloor < s.minNoiseFloor {
+		s.noiseFloor = s.minNoiseFloor
+	}
+}
+
+func (s *Suppressor) applyTransientSuppressionLocked(absSample float32, voicePresence float32, minGain float32, targetGain float32) float32 {
+	s.clickEnergy = (s.clickEnergy * s.clickDecay) + (absSample * (1.0 - s.clickDecay))
+	transient := absSample - s.clickEnergy
+	transientThreshold := 0.04 + (0.08 * (1.0 - voicePresence))
+	if transient > transientThreshold && voicePresence < 0.65 {
+		clickGain := minGain * 0.55
+		if clickGain < targetGain {
+			targetGain = clickGain
+		}
+	}
+	return clampFloat32(targetGain, 0.02, 1.0)
+}
+
+func (s *Suppressor) applyGainSmoothingLocked(targetGain float32) {
+	if targetGain < s.suppressionGain {
+		s.suppressionGain += s.gainAttack * (targetGain - s.suppressionGain)
+	} else {
+		s.suppressionGain += s.gainRelease * (targetGain - s.suppressionGain)
+	}
+	s.suppressionGain = clampFloat32(s.suppressionGain, 0.02, 1.0)
+}
+
+func (s *Suppressor) resetStateLocked() {
+	s.prevInput = 0.0
+	s.prevOutput = 0.0
+	s.envelope = s.minNoiseFloor
+	s.noiseFloor = s.minNoiseFloor
+	s.suppressionGain = 1.0
+	s.clickEnergy = 0.0
+}
+
+func clampFloat32(value float32, min float32, max float32) float32 {
+	if value < min {
+		return min
+	}
+	if value > max {
+		return max
+	}
+	return value
+}
+
 // ProcessSamplesAdvanced applies more sophisticated noise suppression
-// This is a placeholder for future RNNoise integration
+// Placeholder for future RNNoise integration.
 func (s *Suppressor) ProcessSamplesAdvanced(samples []int16) {
-	// TODO: Integrate RNNoise or other advanced algorithms
 	s.ProcessSamples(samples)
-}
+}

noise/suppression_test.go

@@ -0,0 +1,103 @@
+package noise
+
+import (
+	"math"
+	"testing"
+)
+
+func TestSuppressorDisabledBypassesSamples(t *testing.T) {
+	suppressor := NewSuppressor()
+	samples := []int16{100, -200, 300, -400, 500}
+	original := append([]int16(nil), samples...)
+
+	suppressor.ProcessSamples(samples)
+
+	for i := range samples {
+		if samples[i] != original[i] {
+			t.Fatalf("expected sample %d to remain unchanged, got %d want %d", i, samples[i], original[i])
+		}
+	}
+}
+
+func TestSuppressorAttenuatesLowLevelNoise(t *testing.T) {
+	suppressor := NewSuppressor()
+	suppressor.SetEnabled(true)
+	suppressor.SetThreshold(0.08)
+
+	input := makeSineFrame(600, 700)
+	originalRMS := frameRMS(input)
+	processed := append([]int16(nil), input...)
+	suppressor.ProcessSamples(processed)
+	processedRMS := frameRMS(processed)
+
+	if processedRMS >= originalRMS*0.8 {
+		t.Fatalf("expected low-level noise attenuation, got RMS %.2f from %.2f", processedRMS, originalRMS)
+	}
+}
+
+func TestSuppressorPreservesSpeechLikeSignal(t *testing.T) {
+	suppressor := NewSuppressor()
+	suppressor.SetEnabled(true)
+	suppressor.SetThreshold(0.08)
+
+	voice := makeSineFrame(1000, 9000)
+	originalRMS := frameRMS(voice)
+	processed := append([]int16(nil), voice...)
+	suppressor.ProcessSamples(processed)
+	processedRMS := frameRMS(processed)
+
+	if processedRMS <= originalRMS*0.6 {
+		t.Fatalf("expected speech-like signal to be mostly preserved, got RMS %.2f from %.2f", processedRMS, originalRMS)
+	}
+}
+
+func TestHigherThresholdAppliesStrongerSuppression(t *testing.T) {
+	lowSuppressor := NewSuppressor()
+	lowSuppressor.SetEnabled(true)
+	lowSuppressor.SetThreshold(0.02)
+
+	highSuppressor := NewSuppressor()
+	highSuppressor.SetEnabled(true)
+	highSuppressor.SetThreshold(0.20)
+
+	noiseFrame := makeSineFrame(500, 700)
+	lowRMS := runFrameWarmup(lowSuppressor, noiseFrame, 8)
+	highRMS := runFrameWarmup(highSuppressor, noiseFrame, 8)
+
+	if highRMS >= lowRMS*0.8 {
+		t.Fatalf("expected stronger suppression at higher threshold, got low %.2f high %.2f", lowRMS, highRMS)
+	}
+}
+
+func runFrameWarmup(suppressor *Suppressor, frame []int16, repeats int) float64 {
+	var processed []int16
+	for i := 0; i < repeats; i++ {
+		processed = append([]int16(nil), frame...)
+		suppressor.ProcessSamples(processed)
+	}
+	return frameRMS(processed)
+}
+
+func makeSineFrame(frequency float64, amplitude float64) []int16 {
+	const sampleRate = 48000.0
+	const frameSize = 480
+
+	frame := make([]int16, frameSize)
+	for i := 0; i < frameSize; i++ {
+		value := math.Sin((2.0 * math.Pi * frequency * float64(i)) / sampleRate)
+		frame[i] = int16(value * amplitude)
+	}
+	return frame
+}
+
+func frameRMS(samples []int16) float64 {
+	if len(samples) == 0 {
+		return 0.0
+	}
+	var sumSquares float64
+	for _, sample := range samples {
+		normalized := float64(sample) / 32768.0
+		sumSquares += normalized * normalized
+	}
+	return math.Sqrt(sumSquares / float64(len(samples)))
+}