Skip to content

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

tune(asr): add qwen language hint and long-audio segmentation

altic-devca284c606306
diff --git a/Sources/Fluid/Services/QwenAudioProvider.swift b/Sources/Fluid/Services/QwenAudioProvider.swift
index 3d2b1e89..b63a6c85 100644
--- a/Sources/Fluid/Services/QwenAudioProvider.swift
+++ b/Sources/Fluid/Services/QwenAudioProvider.swift
@@ -22,6 +22,8 @@ final class QwenAudioProvider: TranscriptionProvider {
}
let name = "Qwen3 ASR (FluidAudio)"
+ private let sampleRate = 16_000
+ private let maxSegmentSeconds: Int = 25
var isAvailable: Bool {
if #available(macOS 15.0, *) {
@@ -294,14 +296,38 @@ final class QwenAudioProvider: TranscriptionProvider {
}
func transcribeStreaming(_ samples: [Float]) async throws -> ASRTranscriptionResult {
- try await self.transcribeInternal(samples, maxNewTokens: 192)
+ try await self.transcribeInternal(samples, maxNewTokens: 192, languageHint: self.preferredLanguageHint())
}
func transcribeFinal(_ samples: [Float]) async throws -> ASRTranscriptionResult {
- try await self.transcribeInternal(samples, maxNewTokens: 512)
+ let maxSegmentSamples = self.maxSegmentSeconds * self.sampleRate
+ let languageHint = self.preferredLanguageHint()
+
+ if samples.count <= maxSegmentSamples {
+ return try await self.transcribeInternal(samples, maxNewTokens: 512, languageHint: languageHint)
+ }
+
+ var segments: [String] = []
+ var start = 0
+ while start < samples.count {
+ let end = min(start + maxSegmentSamples, samples.count)
+ let segment = Array(samples[start..<end])
+ let result = try await self.transcribeInternal(segment, maxNewTokens: 512, languageHint: languageHint)
+ let text = result.text.trimmingCharacters(in: .whitespacesAndNewlines)
+ if !text.isEmpty {
+ segments.append(text)
+ }
+ start = end
+ }
+
+ return ASRTranscriptionResult(text: segments.joined(separator: " "), confidence: 1.0)
}
- private func transcribeInternal(_ samples: [Float], maxNewTokens: Int) async throws -> ASRTranscriptionResult {
+ private func transcribeInternal(
+ _ samples: [Float],
+ maxNewTokens: Int,
+ languageHint: String?
+ ) async throws -> ASRTranscriptionResult {
if #available(macOS 15.0, *) {
guard let manager = self.manager else {
throw NSError(
@@ -313,7 +339,7 @@ final class QwenAudioProvider: TranscriptionProvider {
let text = try await manager.transcribe(
audioSamples: samples,
- language: String?.none,
+ language: languageHint,
maxNewTokens: maxNewTokens
)
return ASRTranscriptionResult(text: text, confidence: 1.0)
@@ -326,6 +352,17 @@ final class QwenAudioProvider: TranscriptionProvider {
)
}
+ private func preferredLanguageHint() -> String? {
+ guard let preferred = Locale.preferredLanguages.first else { return nil }
+ let code = preferred.split(separator: "-").first?.lowercased() ?? preferred.lowercased()
+ let supported = Set([
+ "zh", "en", "yue", "ar", "de", "fr", "es", "pt", "id", "it", "ko", "ru", "th",
+ "vi", "ja", "tr", "hi", "ms", "nl", "sv", "da", "fi", "pl", "cs", "fil", "fa",
+ "el", "hu", "mk", "ro"
+ ])
+ return supported.contains(code) ? code : nil
+ }
+
func modelsExistOnDisk() -> Bool {
if #available(macOS 15.0, *) {
return Qwen3AsrModels.modelsExist(at: Qwen3AsrModels.defaultCacheDirectory())