grant/ten31-transcripts
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Revert adjacent same-speaker segment collapse

Browse Source 
User found the merged transcript lines harder to read — too many sentences
joined into one statement. Remove SpeakerReconciler.mergeAdjacent, its wiring in
finishBackend (restore the no-LLM early return), and its tests. Back to one
segment per diarized utterance.
This commit is contained in:
Grant Gilliam
2026-06-08 15:52:27 -05:00
parent d9e8c6ceab
commit 5c80e827a1
3 changed files with 8 additions and 70 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Ten31Transcripts/Session/SessionController.swift
+8 -16
View File
@@ -396,32 +396,24 @@ final class SessionController: ObservableObject {
}
}
/// Post-transcription cleanup + LLM passes. Speaker reconciliation (merge split
/// clusters + content-naming) and the readable recap need the gateway LLM; the
/// adjacent-segment collapse does not. So the collapse runs unconditionally and
/// always re-persists `speakers.json`, while the LLM passes are skipped when no
/// model is available. Any failure leaves the last good `speakers.json` intact.
/// Post-transcription LLM passes (best-effort, share one gateway model lookup):
/// reconcile speaker labels (merge split clusters + name from content), then build
/// the readable recap. A missing LLM or any failure leaves speakers.json intact.
private func finishBackend(speakers: SpeakersFile, inputs: ProcessInputs, settings: AppSettings) async {
let llm = GatewayLLMClient(baseURL: settings.backendBaseURL, skipTLS: settings.skipTLSVerification)
let model = await llm.chatModelId() // nil → no LLM on the gateway; LLM passes skipped
guard let model = await llm.chatModelId() else { return } // no LLM on the gateway → skip both
var resolved = speakers
// Reconcile labels (needs the LLM): merge split clusters, dissolve fragments,
// and name placeholders from transcript content.
if let model, settings.reconcileSpeakers, !speakers.segments.isEmpty {
if settings.reconcileSpeakers, !speakers.segments.isEmpty {
self.transcriptStatus = .processing(0, 0)
let fps = RecapEditModel.loadFingerprints(inputs.folder.appendingPathComponent("cluster_fingerprints.json"))
resolved = await SpeakerReconciler.reconcile(file: speakers, fingerprints: fps,
selfName: inputs.selfName, llm: llm, model: model)
try? resolved.write(to: inputs.folder.appendingPathComponent("speakers.json"))
self.transcriptStatus = .done(speakers: resolved.speakers.count, segments: resolved.segments.count)
}
// Collapse adjacent same-speaker segments (no LLM needed) so fragments
// reabsorbed by smoothing read as one clean line, then persist. Always runs
// — even when the LLM is unavailable — so the saved transcript is cleaned up.
resolved = SpeakerReconciler.mergeAdjacent(resolved)
try? resolved.write(to: inputs.folder.appendingPathComponent("speakers.json"))
self.transcriptStatus = .done(speakers: resolved.speakers.count, segments: resolved.segments.count)
guard let model, settings.recapEnabled, !resolved.segments.isEmpty else { return }
guard settings.recapEnabled, !resolved.segments.isEmpty else { return }
let analyzer = RecapAnalyzer(llm: llm, model: model)
guard let result = try? await analyzer.recap(file: resolved, template: settings.defaultTemplate) else { return }
let title = Self.recapTitle(app: inputs.app, sessionId: inputs.sessionId)
Ten31Transcripts/Session/SpeakerReconciler.swift
-22
View File
@@ -82,28 +82,6 @@ enum SpeakerReconciler {
speakers: speakers, segments: result, models: file.models)
}
/// Collapse consecutive segments from the SAME speaker separated by ≤ `maxGap`
/// seconds into one, joining their text — so fragments reabsorbed by smoothing
/// (e.g. "I" then "need to switch it back") read as a single clean line. Pure.
static func mergeAdjacent(_ file: SpeakersFile, maxGap: Double = 2.0) -> SpeakersFile {
let sorted = file.segments.sorted { $0.start < $1.start }
guard !sorted.isEmpty else { return file }
var out: [SpeakersFile.Segment] = []
for s in sorted {
if var last = out.last, last.speaker == s.speaker, s.start - last.end <= maxGap {
let joined = [last.text, s.text].compactMap { $0?.trimmingCharacters(in: .whitespaces) }
.filter { !$0.isEmpty }.joined(separator: " ")
last = .init(start: last.start, end: max(last.end, s.end), speaker: s.speaker,
text: joined.isEmpty ? nil : joined)
out[out.count - 1] = last
} else {
out.append(s)
}
}
return SpeakersFile(sessionId: file.sessionId, app: file.app, durationSec: file.durationSec,
speakers: file.speakers, segments: out, models: file.models)
}
// MARK: - Voiceprint merge (pure)
static func protectedNames(_ file: SpeakersFile, selfName: String) -> Set<String> {