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: harden speaker-labeled file transcription

postoso4e5a89d8f479
diff --git a/Sources/Fluid/Persistence/BackupService.swift b/Sources/Fluid/Persistence/BackupService.swift
index cf3fc71e..1b43ae34 100644
--- a/Sources/Fluid/Persistence/BackupService.swift
+++ b/Sources/Fluid/Persistence/BackupService.swift
@@ -102,6 +102,8 @@ struct SettingsBackupPayload: Codable, Equatable {
let editPromptRoutingScope: SettingsStore.PromptRoutingScope?
let defaultDictationPromptOverride: String?
let defaultEditPromptOverride: String?
+ let fileTranscriptionSpeakerLabelsEnabled: Bool?
+ let fileTranscriptionExpectedSpeakerCount: Int?
}
struct AppBackupDocument: Codable, Equatable {
diff --git a/Sources/Fluid/Persistence/FileTranscriptionHistoryStore.swift b/Sources/Fluid/Persistence/FileTranscriptionHistoryStore.swift
index 3f31fe21..a646acd0 100644
--- a/Sources/Fluid/Persistence/FileTranscriptionHistoryStore.swift
+++ b/Sources/Fluid/Persistence/FileTranscriptionHistoryStore.swift
@@ -18,6 +18,8 @@ struct FileTranscriptionEntry: Codable, Identifiable, Equatable {
let processingTime: TimeInterval
let confidence: Float
let text: String
+ /// Speaker-attributed segments when diarization was enabled; empty otherwise.
+ let speakerSegments: [SpeakerTranscriptSegment]
init(
id: UUID = UUID(),
@@ -26,7 +28,8 @@ struct FileTranscriptionEntry: Codable, Identifiable, Equatable {
duration: TimeInterval,
processingTime: TimeInterval,
confidence: Float,
- text: String
+ text: String,
+ speakerSegments: [SpeakerTranscriptSegment] = []
) {
self.id = id
self.timestamp = timestamp
@@ -35,6 +38,7 @@ struct FileTranscriptionEntry: Codable, Identifiable, Equatable {
self.processingTime = processingTime
self.confidence = confidence
self.text = text
+ self.speakerSegments = speakerSegments
}
init(from result: TranscriptionResult) {
@@ -45,6 +49,38 @@ struct FileTranscriptionEntry: Codable, Identifiable, Equatable {
self.processingTime = result.processingTime
self.confidence = result.confidence
self.text = result.text
+ self.speakerSegments = result.speakerSegments
+ }
+
+ enum CodingKeys: String, CodingKey {
+ case id, timestamp, fileName, duration, processingTime, confidence, text, speakerSegments
+ }
+
+ init(from decoder: Decoder) throws {
+ let c = try decoder.container(keyedBy: CodingKeys.self)
+ self.id = try c.decode(UUID.self, forKey: .id)
+ self.timestamp = try c.decode(Date.self, forKey: .timestamp)
+ self.fileName = try c.decode(String.self, forKey: .fileName)
+ self.duration = try c.decode(TimeInterval.self, forKey: .duration)
+ self.processingTime = try c.decode(TimeInterval.self, forKey: .processingTime)
+ self.confidence = try c.decode(Float.self, forKey: .confidence)
+ self.text = try c.decode(String.self, forKey: .text)
+ // Older history entries predate speaker labels — tolerate a missing key.
+ self.speakerSegments = try c.decodeIfPresent([SpeakerTranscriptSegment].self, forKey: .speakerSegments) ?? []
+ }
+
+ func encode(to encoder: Encoder) throws {
+ var c = encoder.container(keyedBy: CodingKeys.self)
+ try c.encode(self.id, forKey: .id)
+ try c.encode(self.timestamp, forKey: .timestamp)
+ try c.encode(self.fileName, forKey: .fileName)
+ try c.encode(self.duration, forKey: .duration)
+ try c.encode(self.processingTime, forKey: .processingTime)
+ try c.encode(self.confidence, forKey: .confidence)
+ try c.encode(self.text, forKey: .text)
+ if !self.speakerSegments.isEmpty {
+ try c.encode(self.speakerSegments, forKey: .speakerSegments)
+ }
}
/// Preview text for list display (first 80 chars)
@@ -80,7 +116,8 @@ struct FileTranscriptionEntry: Codable, Identifiable, Equatable {
duration: self.duration,
processingTime: self.processingTime,
fileName: self.fileName,
- timestamp: self.timestamp
+ timestamp: self.timestamp,
+ speakerSegments: self.speakerSegments
)
}
}
diff --git a/Sources/Fluid/Persistence/SettingsStore.swift b/Sources/Fluid/Persistence/SettingsStore.swift
index bcecd4df..8fa12707 100644
--- a/Sources/Fluid/Persistence/SettingsStore.swift
+++ b/Sources/Fluid/Persistence/SettingsStore.swift
@@ -3103,7 +3103,9 @@ final class SettingsStore: ObservableObject {
selectedEditPromptID: self.selectedEditPromptID,
editPromptRoutingScope: self.editPromptRoutingScope,
defaultDictationPromptOverride: self.defaultDictationPromptOverride,
- defaultEditPromptOverride: self.defaultEditPromptOverride
+ defaultEditPromptOverride: self.defaultEditPromptOverride,
+ fileTranscriptionSpeakerLabelsEnabled: self.fileTranscriptionSpeakerLabelsEnabled,
+ fileTranscriptionExpectedSpeakerCount: self.fileTranscriptionExpectedSpeakerCount
)
}
@@ -3246,6 +3248,12 @@ final class SettingsStore: ObservableObject {
self.selectedEditPromptID = payload.selectedEditPromptID
self.defaultDictationPromptOverride = payload.defaultDictationPromptOverride
self.defaultEditPromptOverride = payload.defaultEditPromptOverride
+ if let fileTranscriptionSpeakerLabelsEnabled = payload.fileTranscriptionSpeakerLabelsEnabled {
+ self.fileTranscriptionSpeakerLabelsEnabled = fileTranscriptionSpeakerLabelsEnabled
+ }
+ if let fileTranscriptionExpectedSpeakerCount = payload.fileTranscriptionExpectedSpeakerCount {
+ self.fileTranscriptionExpectedSpeakerCount = fileTranscriptionExpectedSpeakerCount
+ }
self.promptModeSelectedPromptID = payload.promptModeSelectedPromptID
self.isSecondaryDictationPromptOff = payload.secondaryDictationPromptOff ?? false
self.normalizePromptSelectionsIfNeeded()
diff --git a/Sources/Fluid/Services/MeetingTranscriptionService.swift b/Sources/Fluid/Services/MeetingTranscriptionService.swift
index 5b2ebb22..0825b376 100644
--- a/Sources/Fluid/Services/MeetingTranscriptionService.swift
+++ b/Sources/Fluid/Services/MeetingTranscriptionService.swift
@@ -529,46 +529,31 @@ final class MeetingTranscriptionService: ObservableObject {
self.currentStatus = "Transcribing speaker segments (\(index + 1)/\(turns.count))..."
self.progress = 0.3 + (Double(index) / Double(turns.count)) * 0.65
- let samples: [Float]
+ let transcribed: (text: String, confidence: Float)?
do {
- samples = try self.readSamples(
- from: audioFile,
- startSeconds: turn.startSeconds,
- endSeconds: turn.endSeconds,
- minimumDurationSeconds: 1.1
- )
- } catch {
- DebugLogger.shared.warning(
- "Skipping speaker segment at \(String(format: "%.1f", turn.startSeconds))s: \(error.localizedDescription)",
- source: "MeetingTranscriptionService"
- )
- continue
- }
-
- // Mirror the standard path's 1-second ASR minimum
- guard samples.count >= 16_000 else { continue }
-
- let chunkResult: ASRTranscriptionResult
- do {
- chunkResult = try await provider.transcribe(samples)
+ transcribed = try await self.transcribeSpeakerTurn(turn, from: audioFile, provider: provider)
} catch {
+ // A genuine audio-read or ASR failure would drop this turn's time range,
+ // leaving a silent gap in the labeled transcript. Abandon the labeled path
+ // so the caller re-transcribes the whole file — baseline behavior is never
+ // at risk (a complete unlabeled transcript beats a labeled one with holes).
DebugLogger.shared.warning(
- "Transcription failed for speaker segment at \(String(format: "%.1f", turn.startSeconds))s: \(error.localizedDescription)",
+ "Speaker labeling aborted at segment \(index + 1)/\(turns.count) (\(String(format: "%.1f", turn.startSeconds))s): \(error.localizedDescription); falling back to standard transcription",
source: "MeetingTranscriptionService"
)
- continue
+ return nil
}
- let text = chunkResult.text.trimmingCharacters(in: .whitespacesAndNewlines)
- guard !text.isEmpty else { continue }
+ // No usable audio in this turn (too short or silent) — skip without losing content.
+ guard let transcribed else { continue }
segments.append(SpeakerTranscriptSegment(
speaker: turn.speakerLabel,
startSeconds: turn.startSeconds,
endSeconds: turn.endSeconds,
- text: text
+ text: transcribed.text
))
- totalConfidence += chunkResult.confidence
+ totalConfidence += transcribed.confidence
}
guard !segments.isEmpty else {
@@ -614,6 +599,59 @@ final class MeetingTranscriptionService: ObservableObject {
return result
}
+ /// Transcribe a single speaker turn, splitting overlong turns into model-sized chunks so a
+ /// long single-speaker stretch never exceeds the ASR input limit (mirrors the standard
+ /// path's 20-minute chunking). Returns the concatenated text and mean confidence, or nil
+ /// when the turn holds no usable audio (too short or silent). Throws on a genuine audio-read
+ /// or ASR failure so the caller can fall back to full-file transcription rather than emit a
+ /// transcript with silent gaps.
+ private func transcribeSpeakerTurn(
+ _ turn: SpeakerDiarizationService.SpeakerTurn,
+ from audioFile: AVAudioFile,
+ provider: TranscriptionProvider
+ ) async throws -> (text: String, confidence: Float)? {
+ // Keep each ASR request well under the model's input limit, matching the standard path.
+ let maxChunkSeconds: Double = 20 * 60
+
+ var ranges: [(start: Double, end: Double)] = []
+ if turn.endSeconds - turn.startSeconds > maxChunkSeconds {
+ var chunkStart = turn.startSeconds
+ while chunkStart < turn.endSeconds {
+ let chunkEnd = min(chunkStart + maxChunkSeconds, turn.endSeconds)
+ ranges.append((chunkStart, chunkEnd))
+ chunkStart = chunkEnd
+ }
+ } else {
+ ranges.append((turn.startSeconds, turn.endSeconds))
+ }
+
+ var pieces: [String] = []
+ var confidenceSum: Float = 0
+ var transcribedChunks = 0
+
+ for range in ranges {
+ let samples = try self.readSamples(
+ from: audioFile,
+ startSeconds: range.start,
+ endSeconds: range.end,
+ minimumDurationSeconds: 1.1
+ )
+ // Mirror the standard path's 1-second ASR minimum.
+ guard samples.count >= 16_000 else { continue }
+
+ let chunkResult = try await provider.transcribe(samples)
+ let text = chunkResult.text.trimmingCharacters(in: .whitespacesAndNewlines)
+ guard !text.isEmpty else { continue }
+
+ pieces.append(text)
+ confidenceSum += chunkResult.confidence
+ transcribedChunks += 1
+ }
+
+ guard transcribedChunks > 0 else { return nil }
+ return (pieces.joined(separator: " "), confidenceSum / Float(transcribedChunks))
+ }
+
/// Read a time range from an audio file as 16kHz mono Float32 samples.
/// Short ranges are widened symmetrically to `minimumDurationSeconds` (clamped to
/// the file bounds) so very brief speaker turns still meet the ASR input minimum.