Audio: speech leveler (auto gain to -20 dBFS speech, -6..+20 dB, gated) between gate and compressor; honest mic meter (drop the +10 dB display boost and x1.2). Creator upload measured -28.5 LUFS; same audio through the new chain: -17.9 LUFS / -0.7 dBFS peak (Unit: audio loudness)

2026-10-09 13:27:53 -07:00
parent 8303b1cf7f
commit b140024ebf
7 changed files with 169 additions and 26 deletions
@@ -16,18 +16,20 @@ public sealed class AudioLevelMeter

    /// <summary>
    /// Maps a linear level (0..1) onto the meter's display scale: -60 dBFS..0 dBFS
-    /// spread linearly across 0..1, with +10 dB of input amplification. Linear RMS
-    /// of real speech or game audio is ~0.01..0.1 (-40..-20 dBFS), which leaves a
-    /// flat (linear) meter looking dead; the log scale makes typical levels occupy
-    /// the bar and the amplification pushes real speech peaks into the red zone
-    /// (0.8+) at maxed volume instead of hovering at its edge. Inputs at or below
-    /// -60 dBFS (0.001 linear) read as zero — the meter never idles on background noise.
+    /// spread linearly across 0..1 — an HONEST dBFS scale. Linear RMS of real speech is
+    /// ~0.01..0.1 (-40..-20 dBFS), which leaves a flat (linear) meter looking dead; the
+    /// log scale makes typical levels occupy the bar. Inputs at or below -60 dBFS
+    /// (0.001 linear) read as zero — the meter never idles on background noise.
+    /// <para>2026-10-09: the +10 dB "input amplification" that used to sit here is gone.
+    /// It made a −30 dBFS voice look dialed in, and a creator's upload came out at
+    /// −28.5 LUFS. The speech leveler now puts voice at −20 dBFS, which lands at 2/3 of
+    /// the bar (the green/yellow edge); loud peaks reach the red.</para>
    /// </summary>
    public static float ToDisplay(float linear)
    {
        if (linear <= 0.001f)
            return 0f;
-        var db = 20f * MathF.Log10(linear) + 10f;
+        var db = 20f * MathF.Log10(linear);
        return Math.Clamp(1f + db / 60f, 0f, 1f);
    }

@@ -3,15 +3,21 @@ namespace ytLive.Services.Audio;
 /// <summary>
 /// The always-on voice processing chain for the mic (TASK 9 audio milestone).
 /// Pure, per-sample, unit-tested: bass boost → treble lift → noise gate →
-/// compressor. Applied BEFORE the meter and the stream mix so what the creator
-/// hears on the meter is what viewers hear — and so background hum is gated
-/// out before it reaches the encoder.
+/// <b>speech leveler</b> → compressor. Applied BEFORE the meter and the stream mix so
+/// what the creator sees on the meter is what viewers hear — and so background hum
+/// is gated out before it reaches the encoder.
+/// <para>The leveler (2026-10-09) is the gain stage the chain was missing: a creator's
+/// upload measured −28.5 LUFS against YouTube's −14 reference because nothing ever
+/// brought quiet speech UP — the compressor only folds peaks above −6 dBFS and the
+/// master limiter only caps at −1 dBFS. Compressor → gain → limiter is the standard
+/// streaming voice chain.</para>
 /// </summary>
 public sealed class VoiceFilterChain
 {
    private readonly LowShelfFilter _bass;
    private readonly HighShelfFilter _treble;
    private readonly NoiseGate _gate;
+    private readonly SpeechLeveler _leveler;
    private readonly Compressor _compressor;

    /// <param name="sampleRate">The (post-resample) mic rate — the RBJ shelf
@@ -21,15 +27,20 @@ public sealed class VoiceFilterChain
        _bass = new LowShelfFilter(sampleRate, 120f, 4f);
        _treble = new HighShelfFilter(sampleRate, 8000f, 3f);
        _gate = new NoiseGate();
+        _leveler = new SpeechLeveler(sampleRate);
        _compressor = new Compressor();
    }

+    /// <summary>The leveler's current gain (linear) — for diagnostics and tests.</summary>
+    public float LevelerGain => _leveler.Gain;
+
    public float Process(float sample)
    {
        var bass = _bass.Process(sample);
        var treble = _treble.Process(bass);
        var gated = _gate.Process(treble);
-        return _compressor.Process(gated);
+        var leveled = _leveler.Process(gated, _gate.IsOpen);
+        return _compressor.Process(leveled);
    }

    public void Reset()
@@ -37,10 +48,87 @@ public sealed class VoiceFilterChain
        _bass.Reset();
        _treble.Reset();
        _gate.Reset();
+        _leveler.Reset();
        _compressor.Reset();
    }
 }

+/// <summary>
+/// Automatic gain for speech: tracks the voice's RMS over ~0.3 s and slowly steers a
+/// gain so speech sits at <see cref="TargetRms"/> (−20 dBFS mono ≈ −17 LUFS once the mic
+/// is mixed to stereo — close to YouTube's −14 reference while leaving the limiter room).
+/// Different mics differ by 20 dB+ in sensitivity; this is what makes "it just sounds
+/// right" true for all of them.
+/// <list type="bullet">
+/// <item>Only learns while the noise gate is open and the voice is above a silence floor —
+/// pauses and background hum never get pumped up.</item>
+/// <item>Gain rises at most <see cref="RiseDbPerSecond"/> and falls at most
+/// <see cref="FallDbPerSecond"/> (faster down than up, so a sudden shout is tamed quickly
+/// and a quiet stretch isn't chased), within <see cref="MinGain"/>..<see cref="MaxGain"/>.</item>
+/// <item>Peaks it pushes up are caught by the <see cref="Compressor"/> after it and the
+/// master limiter at the end of the mix.</item>
+/// </list>
+/// </summary>
+public sealed class SpeechLeveler
+{
+    public const float TargetRms = 0.1f;          // −20 dBFS
+    public const float MaxGain = 10f;             // +20 dB
+    public const float MinGain = 0.5f;            // −6 dB
+    public const float SilenceFloorRms = 0.002f;  // −54 dBFS: below this, hold the gain
+    public const float RiseDbPerSecond = 10f;
+    public const float FallDbPerSecond = 30f;
+    private const float WindowSeconds = 0.3f;
+    private const float UpdateSeconds = 0.01f;
+
+    private readonly float _meanAlpha;
+    private readonly int _updateEvery;
+    private readonly float _maxRiseDb;
+    private readonly float _maxFallDb;
+    private float _meanSquare;
+    private float _gainDb;
+    private float _gain = 1f;
+    private int _sinceUpdate;
+
+    public SpeechLeveler(int sampleRate)
+    {
+        _meanAlpha = 1f / (WindowSeconds * sampleRate);
+        _updateEvery = Math.Max(1, (int)(UpdateSeconds * sampleRate));
+        _maxRiseDb = RiseDbPerSecond * UpdateSeconds;
+        _maxFallDb = FallDbPerSecond * UpdateSeconds;
+    }
+
+    public float Gain => _gain;
+
+    public float Process(float sample, bool voiceActive)
+    {
+        if (voiceActive)
+            _meanSquare += (sample * sample - _meanSquare) * _meanAlpha;
+
+        if (++_sinceUpdate >= _updateEvery)
+        {
+            _sinceUpdate = 0;
+            var rms = MathF.Sqrt(_meanSquare);
+            if (voiceActive && rms > SilenceFloorRms)
+            {
+                var desiredDb = 20f * MathF.Log10(Math.Clamp(TargetRms / rms, MinGain, MaxGain));
+                var step = Math.Clamp(desiredDb - _gainDb, -_maxFallDb, _maxRiseDb);
+                _gainDb += step;
+                _gain = MathF.Pow(10f, _gainDb / 20f);
+            }
+        }
+
+        return sample * _gain;
+    }
+
+    public void Reset()
+    {
+        _meanSquare = 0;
+        _gainDb = 0;
+        _gain = 1f;
+        _sinceUpdate = 0;
+    }
+}
+
 /// <summary>
 /// RBJ audio EQ cookbook low-shelf biquad (bass boost). Transposed direct
 /// form II — stable, cheap, standard.