feat(mac): add Qwen3 MLX streaming dictation

Replace the Sherpa offline pipeline with native MLX streaming, resilient model downloads, live transcript previews, and supporting tests and documentation.
2026-07-23 14:34:56 +08:00
parent f1a811fbf0
commit c0c9dad149
35 changed files with 1373 additions and 764 deletions
@@ -0,0 +1,85 @@
+// MacHallucinationFilter.swift
+// OSGKeyboard · Mac
+//
+// Strips Qwen3 / MLX streaming scaffold tokens and silence hallucinations.
+
+import Foundation
+
+enum MacQwen3LanguageHint {
+    /// Map persisted BCP-47 locale ids to Qwen3 prompt language names.
+    /// Returns `nil` for auto-detect.
+    static func from(locale: Locale) -> String? {
+        let raw = locale.identifier.lowercased()
+        if raw.isEmpty || raw == "auto" { return nil }
+        if raw.hasPrefix("zh") { return "Chinese" }
+        if raw.hasPrefix("en") { return "English" }
+        if raw.hasPrefix("ja") { return "Japanese" }
+        if raw.hasPrefix("ko") { return "Korean" }
+        if raw.hasPrefix("fr") { return "French" }
+        if raw.hasPrefix("de") { return "German" }
+        if raw.hasPrefix("es") { return "Spanish" }
+        if raw.hasPrefix("pt") { return "Portuguese" }
+        if raw.hasPrefix("ru") { return "Russian" }
+        if raw.hasPrefix("ar") { return "Arabic" }
+        return nil
+    }
+}
+
+enum MacHallucinationFilter {
+    /// RMS below this skips feeding audio into the MLX streaming session.
+    static let silencePeakThreshold: Float = 0.0005
+
+    static func strip(_ raw: String) -> String {
+        var text = raw.trimmingCharacters(in: .whitespacesAndNewlines)
+        if text.isEmpty { return "" }
+
+        if let marker = text.range(of: "<asr_text>", options: .backwards) {
+            text = String(text[marker.upperBound...])
+                .trimmingCharacters(in: .whitespacesAndNewlines)
+        } else if let match = text.range(
+            of: #"^language\s+\S+\s*"#,
+            options: [.regularExpression, .caseInsensitive]
+        ) {
+            text = String(text[match.upperBound...])
+                .trimmingCharacters(in: .whitespacesAndNewlines)
+        }
+
+        if isMetadataNoiseLine(text) { return "" }
+        return text
+    }
+
+    /// When the transcript is mostly vocabulary tokens and audio energy stayed low, drop it.
+    static func shouldDiscardHotwordDump(
+        text: String,
+        peakRMS: Float,
+        bias: LocalASRBiasPayload?
+    ) -> Bool {
+        let trimmed = text.trimmingCharacters(in: .whitespacesAndNewlines)
+        guard !trimmed.isEmpty else { return true }
+        guard peakRMS < FlowCaptureTailDrainPolicy.flowDefault.silenceRMSThreshold else {
+            return false
+        }
+        guard let bias, !bias.hardHotwords.isEmpty else { return false }
+        let lowered = trimmed.lowercased()
+        let hits = bias.hardHotwords.filter { lowered.contains($0.lowercased()) }.count
+        let wordCount = max(1, trimmed.split { $0.isWhitespace }.count)
+        return hits >= wordCount
+    }
+
+    private static func isMetadataNoiseLine(_ line: String) -> Bool {
+        let lowered = line.lowercased()
+        switch lowered {
+        case "language", "emotion", "event", "text",
+             "<asr_text>", "</asr_text>", "<|im_end|>":
+            return true
+        default:
+            if lowered.range(
+                of: #"^language(\s+\S+)?$"#,
+                options: .regularExpression
+            ) != nil {
+                return true
+            }
+            return false
+        }
+    }
+}