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

fix cohere long audio duplication

altic-dev76afa265e85f
diff --git a/Sources/Fluid/Services/ExternalCoreMLTranscriptionProvider.swift b/Sources/Fluid/Services/ExternalCoreMLTranscriptionProvider.swift
index b66391d3..5fc50ba3 100644
--- a/Sources/Fluid/Services/ExternalCoreMLTranscriptionProvider.swift
+++ b/Sources/Fluid/Services/ExternalCoreMLTranscriptionProvider.swift
@@ -154,7 +154,10 @@ final class ExternalCoreMLTranscriptionProvider: TranscriptionProvider {
source: "ExternalCoreML"
)
let promptIDs = self.coherePromptIDsForCurrentLanguage()
- let text = try await self.transcribeByManifestWindow(samples, manager: manager, promptIDs: promptIDs)
+ let text = try await manager.transcribe(
+ audioSamples: samples,
+ promptIDs: promptIDs.isEmpty ? nil : promptIDs
+ )
let elapsed = Date().timeIntervalSince(startedAt)
let rtf = audioSeconds > 0 ? elapsed / audioSeconds : 0
DebugLogger.shared.info(
@@ -423,52 +426,6 @@ final class ExternalCoreMLTranscriptionProvider: TranscriptionProvider {
return Array(samples.suffix(maxPreviewSamples))
}
- private func transcribeByManifestWindow(
- _ samples: [Float],
- manager: CohereTranscribeAsrManager,
- promptIDs: [Int]
- ) async throws -> String {
- let runtimePromptIDs = promptIDs.isEmpty ? nil : promptIDs
- let maxAudioSamples = self.loadedManifest?.maxAudioSamples ?? 0
- guard maxAudioSamples > 0 else {
- return try await manager.transcribe(
- audioSamples: samples,
- promptIDs: runtimePromptIDs
- )
- }
-
- if samples.count <= maxAudioSamples {
- return try await manager.transcribe(
- audioSamples: self.paddedSamplesToModelLimit(samples),
- promptIDs: runtimePromptIDs
- )
- }
-
- let overlapSamples = min(self.loadedManifest?.overlapSamples ?? 0, maxAudioSamples / 2)
- let step = max(maxAudioSamples - overlapSamples, 1)
- var chunkTexts: [String] = []
- var startIndex = 0
-
- while startIndex < samples.count {
- let endIndex = min(startIndex + maxAudioSamples, samples.count)
- let chunk = Array(samples[startIndex..<endIndex])
- let text = try await manager.transcribe(
- audioSamples: self.paddedSamplesToModelLimit(chunk),
- promptIDs: runtimePromptIDs
- )
- let trimmed = text.trimmingCharacters(in: .whitespacesAndNewlines)
- if trimmed.isEmpty == false {
- chunkTexts.append(trimmed)
- }
- if endIndex >= samples.count {
- break
- }
- startIndex += step
- }
-
- return chunkTexts.joined(separator: " ")
- }
-
private func paddedSamplesToModelLimit(_ samples: [Float]) -> [Float] {
let maxAudioSamples = self.loadedManifest?.maxAudioSamples ?? samples.count
guard maxAudioSamples > 0 else { return samples }