Repository navigation
Fix: Speak failed cloud-TTS sentences with the Apple voice #64
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -9,6 +9,10 @@ | |
| // | ||
| // The engine is provider-agnostic; a CloudVoiceProvider only says whether it can speak, which | ||
| // voice is selected, and how to turn one sentence into WAV bytes. | ||
| // | ||
| // A sentence whose request fails (a 429, a timeout, no network) is spoken in the Apple voice in | ||
| // its place in the queue, so the reply has no silent hole. After two failures in a row the rest | ||
| // of the reply goes to the Apple voice outright rather than alternating per sentence. | ||
|
|
||
| import AVFoundation | ||
| import Foundation | ||
|
|
@@ -45,23 +49,38 @@ final class CloudTTSService: ObservableObject { | |
| /// The format the player node is connected with (the first clip's, normally 24 kHz mono). | ||
| private var playerFormat: AVAudioFormat? | ||
|
|
||
| /// Speaks failed sentences. Its own synthesizer, not TTSService: the voice agent reads | ||
| /// TTSService finishing as the end of the reply. | ||
| private let fallback = AppleFallbackSpeaker() | ||
| /// Whether the current utterance falls back to the Apple voice (replies yes, voice previews | ||
| /// no: a failed preview shows `lastFailure` instead). | ||
| private var fallbackEnabled = true | ||
| /// Failed sentences in a row; at two, the rest of the reply uses the Apple voice. | ||
| private var consecutiveFailures = 0 | ||
| private var appleForRest = false | ||
| /// Test hook: each sentence as it starts playing, and whether it's in the Apple voice. | ||
| var onSentenceStarted: ((_ sentence: String, _ appleVoice: Bool) -> Void)? | ||
| /// Test hook: how many sentences the Apple voice has actually been asked to speak. | ||
| var fallbackUtterances: Int { fallback.utterances } | ||
|
|
||
| init(provider: CloudVoiceProvider) { | ||
| self.provider = provider | ||
| } | ||
|
|
||
| // MARK: - Speak | ||
|
|
||
| /// Speak a whole reply, replacing anything playing. | ||
| func speak(_ text: String) async { | ||
| await speak(text, voice: provider.selectedVoice(), appleFallback: true) | ||
| } | ||
|
|
||
| /// Speak in a specific voice — used to preview voices in Settings before choosing one. | ||
| func speak(_ text: String, voice: String, appleFallback: Bool = false) async { | ||
| let clean = text.trimmingCharacters(in: .whitespacesAndNewlines) | ||
| guard !clean.isEmpty, isReady else { return } | ||
| stop() | ||
| beginStreaming() | ||
| fallbackEnabled = appleFallback | ||
| for sentence in TextChunking.sentences(clean) { enqueue(sentence, voice: voice) } | ||
| endStreaming() | ||
| } | ||
|
|
@@ -72,11 +91,12 @@ final class CloudTTSService: ObservableObject { | |
| streamStarted = false | ||
| utteranceCancelled = true // in-flight synthesis discards its output | ||
| pendingSentences.removeAll() | ||
| clipQueue.forEach { $0.clip.cancel() } | ||
| clipQueue.removeAll() | ||
| pendingBuffers = 0 | ||
| playerEpoch += 1 | ||
| if audioReady { playerNode.stop() } | ||
| fallback.stop() | ||
| } | ||
|
|
||
| // MARK: - Streaming (speak sentences as the model produces them) | ||
|
|
@@ -88,7 +108,7 @@ final class CloudTTSService: ObservableObject { | |
| /// Synthesis tasks in sentence order, at most `maxInFlight` at a time (enough to stay ahead of | ||
| /// playback without a 15-sentence reply firing 15 requests and tripping rate limits). ONE | ||
| /// drainer awaits them in order, so playback order is structural. | ||
| private var clipQueue: [(text: String, clip: Task<AVAudioPCMBuffer?, Never>)] = [] | ||
| /// Sentences waiting for a synthesis slot. | ||
| private var pendingSentences: [(text: String, voice: String)] = [] | ||
| private let maxInFlight = 3 | ||
|
|
@@ -123,6 +143,9 @@ final class CloudTTSService: ObservableObject { | |
| utteranceCancelled = false | ||
| generationActive = true | ||
| utteranceGeneration += 1 | ||
| fallbackEnabled = true | ||
| consecutiveFailures = 0 | ||
| appleForRest = false | ||
| isSpeaking = true // pauses the recognizer for the whole utterance | ||
| } | ||
|
|
||
|
|
@@ -142,7 +165,9 @@ final class CloudTTSService: ObservableObject { | |
| private func startSynthesis() { | ||
| while clipQueue.count < maxInFlight, !pendingSentences.isEmpty { | ||
| let next = pendingSentences.removeFirst() | ||
| // Once the reply has moved to the Apple voice, don't request it from the cloud. | ||
| let skip = appleForRest | ||
| clipQueue.append((next.text, Task { skip ? nil : await self.synthesize(next.text, voice: next.voice) })) | ||
| } | ||
| } | ||
|
|
||
|
|
@@ -155,16 +180,24 @@ final class CloudTTSService: ObservableObject { | |
| private func drainIfNeeded() { | ||
| guard drainTask == nil else { return } | ||
| drainTask = Task { [weak self] in | ||
| while let self, let head = self.clipQueue.first { | ||
| let clip = await head.clip.value | ||
| // stop() (and maybe a new reply) may have replaced the queue while we waited — | ||
| // only play the clip if it's still the head of the CURRENT utterance. | ||
| guard self.clipQueue.first?.clip == head.clip, !self.utteranceCancelled else { continue } | ||
| self.clipQueue.removeFirst() | ||
| let useCloud = clip != nil && !self.appleForRest | ||
| // Count the failure before refilling the queue, so a second failure in a row | ||
| // keeps the next sentence from being sent to the cloud. | ||
| if useCloud { self.consecutiveFailures = 0 } else if self.fallbackEnabled { self.noteFailure() } | ||
| self.startSynthesis() | ||
| if let clip, useCloud { | ||
| self.onSentenceStarted?(head.text, false) | ||
| guard !self.utteranceCancelled else { continue } | ||
| self.schedule(clip, restartPlayer: !self.streamStarted) | ||
| self.streamStarted = true | ||
| } else if self.fallbackEnabled { | ||
| await self.speakWithAppleVoice(head.text) | ||
| } | ||
| } | ||
| guard let self else { return } | ||
|
|
@@ -173,6 +206,32 @@ final class CloudTTSService: ObservableObject { | |
| } | ||
| } | ||
|
|
||
| /// Speak a sentence the cloud couldn't synthesize, in its place: after the clips queued before | ||
| /// it have played out, and before the next one. The drainer waits here, so the two audio | ||
| /// paths never overlap and `isSpeaking` stays true throughout. | ||
| private func speakWithAppleVoice(_ sentence: String) async { | ||
| // A stop, then a new reply, can happen during the wait: only speak for this one. | ||
| let generation = utteranceGeneration | ||
| var current: Bool { !utteranceCancelled && utteranceGeneration == generation } | ||
| while pendingBuffers > 0 && current { | ||
| try? await Task.sleep(for: .milliseconds(30)) | ||
| } | ||
| guard current else { return } | ||
| onSentenceStarted?(sentence, true) | ||
| guard current else { return } | ||
| MetricsCollector.shared.markFirstAudio() | ||
| await fallback.speak(sentence) | ||
| } | ||
|
|
||
| private func noteFailure() { | ||
| consecutiveFailures += 1 | ||
| guard consecutiveFailures >= 2, !appleForRest else { return } | ||
| // Probably offline: stop asking the cloud for the rest of this reply. | ||
| NSLog("[%@] two sentences failed in a row, using the Apple voice for the rest", provider.name) | ||
| appleForRest = true | ||
| clipQueue.forEach { $0.clip.cancel() } // requests already out come back as failures | ||
| } | ||
|
|
||
| /// Ambient narration (watch loop): speaks only into silence and drops itself if a reply | ||
| /// started while it was being synthesized. | ||
| func speakAmbient(_ text: String) async { | ||
|
|
@@ -286,6 +345,70 @@ final class CloudTTSService: ObservableObject { | |
|
|
||
| extension CloudTTSService: NeuralSpeechEngine {} | ||
|
|
||
| // MARK: - Apple voice fallback | ||
|
|
||
| /// Speaks one sentence in the user's Apple voice and returns when it has finished or been stopped. | ||
| @MainActor | ||
| final class AppleFallbackSpeaker: NSObject, AVSpeechSynthesizerDelegate { | ||
| private let synthesizer = AVSpeechSynthesizer() | ||
| /// The utterance being spoken and its waiter. A stopped utterance reports didCancel later, | ||
| /// so callbacks only finish the utterance they're about. | ||
| private var current: (id: ObjectIdentifier, finished: CheckedContinuation<Void, Never>, watchdog: Task<Void, Never>)? | ||
| private(set) var utterances = 0 | ||
|
|
||
| /// How long one sentence may take before we stop waiting for the synthesizer: the drainer | ||
| /// waits here with the recognizer paused, so a lost didFinish must not leave it deaf. | ||
| nonisolated static func deadline(for sentence: String) -> Duration { | ||
| .milliseconds(max(5_000, 120 * sentence.count)) | ||
| } | ||
|
|
||
| override init() { | ||
| super.init() | ||
| synthesizer.delegate = self | ||
| } | ||
|
|
||
| func speak(_ sentence: String) async { | ||
| stop() | ||
| let utterance = AVSpeechUtterance(string: sentence) | ||
| utterance.voice = TTSService.userVoice | ||
| utterances += 1 | ||
| let id = ObjectIdentifier(utterance) | ||
| await withCheckedContinuation { continuation in | ||
|
Owner
There was a problem hiding this comment. The cloud path bounds its wait with Please cap it, e.g. start a
Contributor
Author
There was a problem hiding this comment. Fixed in 67f938e, as you suggested. |
||
| let watchdog = Task { [weak self] in | ||
| try? await Task.sleep(for: Self.deadline(for: sentence)) | ||
| guard !Task.isCancelled, let self, self.current?.id == id else { return } | ||
| NSLog("[CloudTTS] Apple voice never finished a sentence; moving on") | ||
| self.synthesizer.stopSpeaking(at: .immediate) | ||
| self.finish(id) | ||
| } | ||
| current = (id, continuation, watchdog) | ||
| synthesizer.speak(utterance) | ||
| } | ||
| } | ||
|
|
||
| func stop() { | ||
| synthesizer.stopSpeaking(at: .immediate) | ||
| if let current { finish(current.id) } | ||
| } | ||
|
|
||
| private func finish(_ id: ObjectIdentifier) { | ||
| guard let current, current.id == id else { return } | ||
| self.current = nil | ||
| current.watchdog.cancel() | ||
| current.finished.resume() | ||
| } | ||
|
|
||
| nonisolated func speechSynthesizer(_ synthesizer: AVSpeechSynthesizer, didFinish utterance: AVSpeechUtterance) { | ||
| let id = ObjectIdentifier(utterance) | ||
| Task { @MainActor in self.finish(id) } | ||
| } | ||
|
|
||
| nonisolated func speechSynthesizer(_ synthesizer: AVSpeechSynthesizer, didCancel utterance: AVSpeechUtterance) { | ||
| let id = ObjectIdentifier(utterance) | ||
| Task { @MainActor in self.finish(id) } | ||
| } | ||
| } | ||
|
|
||
| // MARK: - Voices | ||
|
|
||
| struct CloudVoice: Identifiable, Hashable { | ||
|
|
||
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Uh oh!
There was an error while loading. Please reload this page.