From 78caac8d0460f71fef360f3be0b4e50e00c629ea Mon Sep 17 00:00:00 2001
From: hariel1985 <ariel.hartung@gmail.com>
Date: Wed, 16 Sep 2026 12:25:58 +0200
Subject: [PATCH] Fix silent live preview on stereo and multi-input audio
 interfaces

AVAudioConverter keeps only channel 1 when downmixing stereo and outputs
pure silence for layout-less multi-channel inputs (e.g. a 4-input USB
interface such as the UA Volt 476P), so the dictation model heard nothing
and the HUD never showed text. Pick the channel that carries the voice
(loudest on smoothed energy, switching only on a clear 6 dB margin) and
convert it as mono. Also reserve the dictation locale, which the Speech
framework warns will become required, and log the mic format.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
---
 macos/src/main.swift | 54 ++++++++++++++++++++++++++++++++++++++++----
 1 file changed, 50 insertions(+), 4 deletions(-)

diff --git a/macos/src/main.swift b/macos/src/main.swift
index dd2937d..da969c7 100644
--- a/macos/src/main.swift
+++ b/macos/src/main.swift
@@ -468,6 +468,7 @@ final class LivePreview {
                 let transcriber = LivePreview.makeTranscriber(locale: locale)
                 // Even when the model is already on disk the app may first have to register it;
                 // that takes a fraction of a second. A real download just means no preview this time.
+                _ = try? await AssetInventory.reserve(locale: locale)
                 if await AssetInventory.status(forModules: [transcriber]) != .installed,
                    let request = try await AssetInventory.assetInstallationRequest(supporting: [transcriber]) {
                     NSLog("Live preview: preparing dictation model for \(locale.identifier)")
@@ -482,7 +483,9 @@ final class LivePreview {
                 let engine = AVAudioEngine()
                 let micFormat = engine.inputNode.outputFormat(forBus: 0)
                 guard micFormat.sampleRate > 0, current == self.session,
-                      let converter = AVAudioConverter(from: micFormat, to: analyzerFormat) else { return }
+                      let monoFormat = AVAudioFormat(standardFormatWithSampleRate: micFormat.sampleRate, channels: 1),
+                      let converter = AVAudioConverter(from: monoFormat, to: analyzerFormat) else { return }
+                NSLog("Live preview: mic %d ch @ %.0f Hz -> %@", micFormat.channelCount, micFormat.sampleRate, analyzerFormat.description)
                 engine.inputNode.installTap(onBus: 0, bufferSize: 4096, format: micFormat,
                                             block: LivePreview.tapBlock(converter: converter, format: analyzerFormat, continuation: continuation))
                 try engine.start()
@@ -543,11 +546,13 @@ final class LivePreview {
         return head + " " + tail
     }
 
-    // Runs on the audio thread: convert mic buffers to the analyzer's format.
-    // Built in a nonisolated static context so it carries no actor assumptions.
+    // Runs on the audio thread: take the voice channel as mono, convert it to the
+    // analyzer's format. Built in a static context so it carries no actor assumptions.
     private static func tapBlock(converter: AVAudioConverter, format: AVAudioFormat,
                                  continuation: AsyncStream<AnalyzerInput>.Continuation) -> AVAudioNodeTapBlock {
-        return { buffer, _ in
+        let downmixer = VoiceChannelPicker()
+        return { micBuffer, _ in
+            guard let buffer = downmixer.mono(from: micBuffer) else { return }
             let ratio = format.sampleRate / buffer.format.sampleRate
             let capacity = AVAudioFrameCount(Double(buffer.frameLength) * ratio) + 64
             guard let out = AVAudioPCMBuffer(pcmFormat: format, frameCapacity: capacity) else { return }
@@ -569,6 +574,47 @@ final class LivePreview {
     }
 }
 
+// Picks the input channel that carries the voice and returns it as a mono buffer.
+// AVAudioConverter keeps only channel 1 when downmixing stereo, and outputs pure
+// silence for layout-less multi-channel inputs (e.g. a 4-input USB interface),
+// so the live preview model would hear nothing. The loudest channel is chosen on
+// smoothed energy, and only replaced by one that is clearly (6 dB) louder.
+final class VoiceChannelPicker {
+    private var energy: [Float] = []
+    private var active = 0
+    private var monoFormat: AVAudioFormat?
+
+    func mono(from buffer: AVAudioPCMBuffer) -> AVAudioPCMBuffer? {
+        let channels = Int(buffer.format.channelCount)
+        let frames = Int(buffer.frameLength)
+        guard frames > 0, let data = buffer.floatChannelData else { return nil }
+        if channels == 1 { return buffer }
+
+        if energy.count != channels {
+            energy = Array(repeating: 0, count: channels)
+            active = 0
+        }
+        for c in 0..<channels {
+            var sum: Float = 0
+            for i in 0..<frames { sum += data[c][i] * data[c][i] }
+            energy[c] = energy[c] * 0.9 + sum / Float(frames) * 0.1
+        }
+        if let loudest = energy.indices.max(by: { energy[$0] < energy[$1] }), energy[loudest] > energy[active] * 4 {
+            active = loudest
+        }
+
+        if monoFormat?.sampleRate != buffer.format.sampleRate {
+            monoFormat = AVAudioFormat(standardFormatWithSampleRate: buffer.format.sampleRate, channels: 1)
+        }
+        guard let format = monoFormat,
+              let out = AVAudioPCMBuffer(pcmFormat: format, frameCapacity: AVAudioFrameCount(frames)),
+              let dest = out.floatChannelData else { return nil }
+        dest[0].update(from: data[active], count: frames)
+        out.frameLength = AVAudioFrameCount(frames)
+        return out
+    }
+}
+
 final class MeterView: NSView {
     enum Mode { case recording, transcribing }
     enum Notice { case loadingModel, modelLoaded }