Fastest and only macOS Dictation app with on-device STT and custom trained AI enhancement model. Windows, iOS testflight! Android, Linux waitlist open! A local Wispr Flow alternative. Follow https://x.com/fluidvoiceapp for more
fix(asr): reject sliding final artifacts
altic-devec4bf1fab2a5
diff --git a/Sources/Fluid/Services/FluidAudioProvider.swift b/Sources/Fluid/Services/FluidAudioProvider.swiftindex bd81aef2..1383dd1c 100644--- a/Sources/Fluid/Services/FluidAudioProvider.swift+++ b/Sources/Fluid/Services/FluidAudioProvider.swift@@ -307,6 +307,14 @@ final class FluidAudioProvider: TranscriptionProvider {DebugLogger.shared.info("ASR_BENCH provider_sliding_rejected reason=mid_word_period", source: "ASRBenchmark")return false}+ guard !self.hasSuspiciousAdjacentPunctuation(cleaned) else {+ DebugLogger.shared.info("ASR_BENCH provider_sliding_rejected reason=adjacent_punctuation", source: "ASRBenchmark")+ return false+ }+ guard !self.hasRepeatedAdjacentPhrase(cleaned) else {+ DebugLogger.shared.info("ASR_BENCH provider_sliding_rejected reason=repeated_phrase", source: "ASRBenchmark")+ return false+ }guard !self.latestStreamingPreviewText.isEmpty else { return true }let minimumLength = Int(Double(self.latestStreamingPreviewText.count) * 0.92)@@ -331,6 +339,32 @@ final class FluidAudioProvider: TranscriptionProvider {return false}+ private func hasSuspiciousAdjacentPunctuation(_ text: String) -> Bool {+ text.contains(".,") || text.contains(",.") || text.contains("..")+ }++ private func hasRepeatedAdjacentPhrase(_ text: String) -> Bool {+ let words = text+ .lowercased()+ .split { !$0.isLetter && !$0.isNumber }+ .map(String.init)+ guard words.count >= 4 else { return false }++ let maxPhraseLength = min(5, words.count / 2)+ for phraseLength in 2...maxPhraseLength {+ let maxStart = words.count - phraseLength * 2+ guard maxStart >= 0 else { continue }+ for start in 0...maxStart {+ let first = words[start..<(start + phraseLength)]+ let second = words[(start + phraseLength)..<(start + phraseLength * 2)]+ if first.elementsEqual(second) {+ return true+ }+ }+ }+ return false+ }+private func createPCMBuffer(from samples: [Float]) throws -> AVAudioPCMBuffer {guardlet format = AVAudioFormat(commonFormat: .pcmFormatFloat32, sampleRate: 16_000, channels: 1, interleaved: false),