Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Original file line number Diff line number Diff line change
Expand Up @@ -58,6 +58,13 @@ internal class QueuedTtsPlayback(
if (audioFile == null) {
request.completion.complete(false)
} else {
// 合成完成即放行调用方(complete 必须在 send 之前):调用方拿到结果后
// 就能提交下一句,当前句播放期间下一句照常预合成,流水线得以保持。
// 若等播放结束才返回,调用方(ChatViewModel 的串行消费者)每句都要
// 等上一句播完才发起合成,句间会出现可感知的空窗。
// 播放顺序仍由 playbackQueue 串行保证,打断/停止语义由
// generation 与 isCurrent 校验保证,均不受影响。
request.completion.complete(true)
playbackQueue.send(PreparedRequest(request, audioFile))
}
} catch (e: Exception) {
Expand All @@ -69,15 +76,20 @@ internal class QueuedTtsPlayback(
scope.launch {
for (prepared in playbackQueue) {
try {
val result = playPreparedRequest(prepared)
prepared.request.completion.complete(result)
// 调用方在合成完成时已拿到结果,播放阶段的失败(文件损坏等,
// playAudioFile 内部已记录日志)只影响本句发声,不再回传。
playPreparedRequest(prepared)
} catch (e: Exception) {
prepared.request.completion.completeExceptionally(e)
AppLogger.e(tag, "TTS播放队列处理失败", e)
}
}
}
}

/**
* 提交一段文本到播放队列。返回 true 表示合成成功且已进入播放队列(不等待播放结束),
* false 表示合成失败。播放由 playbackQueue 串行推进,打断/停止通过 generation 校验生效。
*/
suspend fun speak(
text: String,
interrupt: Boolean,
Expand Down Expand Up @@ -204,12 +216,11 @@ internal class QueuedTtsPlayback(
}
}

private suspend fun playPreparedRequest(prepared: PreparedRequest): Boolean {
private suspend fun playPreparedRequest(prepared: PreparedRequest) {
if (!isCurrent(prepared.request)) {
return false
return
}
playAudioFile(prepared.audioFile)
return true
}

private suspend fun playAudioFile(audioFile: File) {
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -116,6 +116,13 @@ class SiliconFlowVoiceProvider(
if (prepared == null) {
request.completion.complete(false)
} else {
// 合成完成即放行调用方(complete 必须在 send 之前):调用方拿到结果后
// 就能提交下一句,当前句播放期间下一句照常预合成,流水线得以保持。
// 若等播放结束才返回,调用方(ChatViewModel 的串行消费者)每句都要
// 等上一句播完才发起合成,句间会出现可感知的空窗。
// 播放顺序仍由 playbackQueue 串行保证,打断/停止语义由
// generation 与 isCurrent 校验保证,均不受影响。
request.completion.complete(true)
playbackQueue.send(prepared)
}
} catch (e: Exception) {
Expand All @@ -127,10 +134,11 @@ class SiliconFlowVoiceProvider(
speakScope.launch {
for (prepared in playbackQueue) {
try {
val result = playPreparedRequest(prepared)
prepared.request.completion.complete(result)
// 调用方在合成完成时已拿到结果,播放阶段的失败(文件损坏等,
// playAudioFileAndAwait 内部已记录日志)只影响本句发声,不再回传。
playPreparedRequest(prepared)
} catch (e: Exception) {
prepared.request.completion.completeExceptionally(e)
AppLogger.e(TAG, "TTS播放队列处理失败", e)
}
}
}
Expand All @@ -156,6 +164,10 @@ class SiliconFlowVoiceProvider(
}
}

/**
* 提交一段文本到播放队列。返回 true 表示合成成功且已进入播放队列(不等待播放结束),
* false 表示合成失败。播放由 playbackQueue 串行推进,打断/停止通过 generation 校验生效。
*/
override suspend fun speak(
text: String,
interrupt: Boolean,
Expand Down Expand Up @@ -314,13 +326,13 @@ class SiliconFlowVoiceProvider(
}
}

private suspend fun playPreparedRequest(prepared: PreparedRequest): Boolean {
private suspend fun playPreparedRequest(prepared: PreparedRequest) {
if (prepared.request.generation != stopGeneration.get()) {
prepared.audioFile.delete()
return false
return
}

return playAudioFileAndAwait(prepared.audioFile)
playAudioFileAndAwait(prepared.audioFile)
}

private fun clearPendingRequests() {
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -39,9 +39,11 @@ import com.ai.assistance.operit.ui.floating.FloatingMode
import com.ai.assistance.operit.ui.permissions.PermissionLevel
import com.ai.assistance.operit.ui.permissions.ToolPermissionSystem
import java.io.IOException
import java.util.concurrent.atomic.AtomicLong
import kotlinx.coroutines.Dispatchers
import kotlinx.coroutines.Job
import kotlinx.coroutines.CancellationException
import kotlinx.coroutines.channels.Channel
import kotlinx.coroutines.delay
import kotlinx.coroutines.flow.filter
import kotlinx.coroutines.flow.MutableStateFlow
Expand Down Expand Up @@ -136,6 +138,24 @@ class ChatViewModel(private val context: Context) : ViewModel() {
private var voiceStateCollectionJob: Job? = null
private var speechPlaybackJob: Job? = null
private var speechControlsHideJob: Job? = null

// 朗读请求串行化:所有 speakMessage 调用先进队,由单一消费者协程按入队顺序逐条提交给语音服务。
// 必须串行的原因:provider.speak 的提交顺序即播放顺序(系统 TTS 引擎队列、队列型 Provider 的
// speakQueue 都按提交顺序播放),此前自动朗读每次 flush 起一个并发协程,提交顺序取决于线程
// 调度,句子会乱序进入播放端(自动朗读跳句的根因)。请求到达 Provider 时 flush 顺序信息已经
// 丢失,Provider 侧加锁无法恢复顺序,所以必须在调用方收敛为单点串行。
private data class SpeakRequest(
val message: String,
val interrupt: Boolean,
val generation: Long,
)

private val speakRequests = Channel<SpeakRequest>(Channel.UNLIMITED)

// 代数用于打断/停止后拒绝残留请求:cancel 与 drain 覆盖不到"消费者刚取出、还没开始处理"
// 的窗口,靠代数比对丢弃,避免停止或打断后旧内容漏播一句。与 QueuedTtsPlayback 的
// stopGeneration 是同一套模式。
private val speakRequestGeneration = AtomicLong(0)
private val speechServiceProfiles = SpeechServiceProfilesPreferences(context)
private val activePromptManager = ActivePromptManager.getInstance(context)
private val characterCardManager = CharacterCardManager.getInstance(context)
Expand Down Expand Up @@ -425,6 +445,9 @@ class ChatViewModel(private val context: Context) : ViewModel() {
// 初始化语音服务
initializeVoiceService()

// 启动朗读请求串行消费者,生命周期跟随 viewModelScope
startSpeakRequestConsumer()

// 配置提示只跟随当前活跃聊天配置,避免被其他未使用配置误伤。
viewModelScope.launch {
combine(
Expand Down Expand Up @@ -3019,81 +3042,139 @@ class ChatViewModel(private val context: Context) : ViewModel() {

fun speakMessage(message: String, interrupt: Boolean) {
if (interrupt) {
// 打断三件事:提升代数、取消当前播放、清空积压请求。
// 代数与清空共同保证上一轮残留的 flush(已入队未处理的)不会插播到新一轮朗读中间,
// 与系统 TTS 的 QUEUE_FLUSH / 队列型 Provider 的 clearForInterrupt 语义对齐。
speakRequestGeneration.incrementAndGet()
speechPlaybackJob?.cancel()
drainSpeakRequests("interrupt")
logSpeechState("cancelPlaybackJob", "reason=newInterruptingSpeak")
}
speechPlaybackJob = viewModelScope.launch {
try {
cancelSpeechControlsHide("newSpeak")
val currentVoiceService = ensureActiveVoiceService()
_isSpeechSessionActive.value = true
_isSpeechPaused.value = false
logSpeechState(
"speakMessage.start",
"interrupt=$interrupt provider=${currentVoiceService?.javaClass?.simpleName} rawLen=${message.length} preview=\"${speechPreview(message)}\""
)
if (currentVoiceService == null) {
_isSpeechSessionActive.value = false
logSpeechState("speakMessage.abort", "reason=noVoiceService")
uiStateDelegate.showToast(context.getString(R.string.chat_voice_service_init_failed))
AppLogger.e(TAG, "语音服务不可用")
return@launch
}
speakRequests.trySend(
SpeakRequest(
message = message,
interrupt = interrupt,
generation = speakRequestGeneration.get()
)
)
}

val cleanerRegexs = speechServiceProfiles.getCurrentTtsProfile().cleanerRegexs
val cleanedText = TtsCleaner.cleanForSpeech(message, cleanerRegexs)
val cleanMessage = WaifuMessageProcessor.cleanContentForWaifu(cleanedText)
AppLogger.d(
TAG,
"speech[cleaned] rawLen=${message.length} cleanedLen=${cleanedText.length} finalLen=${cleanMessage.length} preview=\"${speechPreview(cleanMessage)}\""
)
private fun drainSpeakRequests(reason: String) {
while (true) {
val dropped = speakRequests.tryReceive().getOrNull() ?: break
logSpeechState("speakRequest.dropped", "reason=$reason rawLen=${dropped.message.length}")
}
}

if (cleanMessage.isBlank()) {
AppLogger.d(TAG, "朗读内容为空,跳过请求")
_isSpeechSessionActive.value = false
logSpeechState("speakMessage.abort", "reason=blankAfterClean")
return@launch
/**
* 朗读请求的单一串行消费者。手动整条、自动流式逐句、waifu 分段全部经由
* speakMessage 入队,在这里严格按入队顺序处理:处理完一条(子协程结束)才取下一条,
* 保证 provider.speak 的调用顺序与请求顺序一致。
* 不串行的后果:并发协程的提交顺序取决于线程调度,句子乱序进入播放端(系统 TTS 跳句)。
* 队列型 Provider 内部自带预合成流水线(QueuedTtsPlayback 的 prepare 循环),
* 这里的串行等待不会造成句间空窗。
*/
private fun startSpeakRequestConsumer() {
viewModelScope.launch {
for (request in speakRequests) {
if (request.generation != speakRequestGeneration.get()) {
// 打断/停止之后才被消费者取出的残留请求,直接丢弃
logSpeechState("speakRequest.stale", "generation=${request.generation} rawLen=${request.message.length}")
continue
}
val playbackJob = launch {
processSpeakRequest(request)
}
speechPlaybackJob = playbackJob
playbackJob.join()
}
}
}

val segments = TtsSegmenter.split(cleanMessage)
AppLogger.d(TAG, "speech[segments] count=${segments.size} lengths=${segments.joinToString(prefix = "[", postfix = "]") { it.length.toString() }}")
var isFirstSegment = true
for ((index, segment) in segments.withIndex()) {
if (_isSpeechPaused.value) {
logSpeechState("waitResume", "segmentIndex=$index")
isSpeechPaused.filter { !it }.first()
logSpeechState("resumeObserved", "segmentIndex=$index")
}
private suspend fun processSpeakRequest(request: SpeakRequest) {
// Main.immediate 下子协程体会在 launch 返回前内联启动,消费者循环里的代数校验
// 覆盖不到"校验通过到子协程真正开始"的窗口,这里再校验一次(与 QueuedTtsPlayback
// 在合成、播放两处校验 isCurrent 是同一模式)。
if (request.generation != speakRequestGeneration.get()) {
logSpeechState("speakRequest.stale", "generation=${request.generation} rawLen=${request.message.length}")
return
}
val message = request.message
val interrupt = request.interrupt
try {
cancelSpeechControlsHide("newSpeak")
val currentVoiceService = ensureActiveVoiceService()
_isSpeechSessionActive.value = true
_isSpeechPaused.value = false
logSpeechState(
"speakMessage.start",
"interrupt=$interrupt provider=${currentVoiceService?.javaClass?.simpleName} rawLen=${message.length} preview=\"${speechPreview(message)}\""
)
if (currentVoiceService == null) {
_isSpeechSessionActive.value = false
logSpeechState("speakMessage.abort", "reason=noVoiceService")
uiStateDelegate.showToast(context.getString(R.string.chat_voice_service_init_failed))
AppLogger.e(TAG, "语音服务不可用")
return
}

val cleanerRegexs = speechServiceProfiles.getCurrentTtsProfile().cleanerRegexs
// 清理链沿用仓库既有实现:cleanForSpeech 先剥离 provider 协议 meta、思考链与
// tool/status/emotion 块(extractReplyBody),再套用用户 cleaner 规则,
// 最后交给 cleanContentForWaifu。串行化改造不复制这套链,避免同一内容被
// 两套过滤器重复裁剪。
val cleanedText = TtsCleaner.cleanForSpeech(message, cleanerRegexs)
val cleanMessage = WaifuMessageProcessor.cleanContentForWaifu(cleanedText)
AppLogger.d(
TAG,
"speech[cleaned] rawLen=${message.length} cleanedLen=${cleanedText.length} finalLen=${cleanMessage.length} preview=\"${speechPreview(cleanMessage)}\""
)

AppLogger.d(
TAG,
"speech[segmentSpeak] index=$index/${segments.lastIndex} interrupt=${if (isFirstSegment) interrupt else false} len=${segment.length} preview=\"${speechPreview(segment)}\""
)
val success = currentVoiceService.speak(
text = segment,
interrupt = if (isFirstSegment) interrupt else false,
rate = null,
pitch = null
)
AppLogger.d(TAG, "speech[segmentResult] index=$index success=$success provider=${currentVoiceService.javaClass.simpleName}")
if (cleanMessage.isBlank()) {
AppLogger.d(TAG, "朗读内容为空,跳过请求")
_isSpeechSessionActive.value = false
logSpeechState("speakMessage.abort", "reason=blankAfterClean")
return
}

if (!success) {
logSpeechState("segmentFailed", "index=$index")
uiStateDelegate.showToast(context.getString(R.string.chat_speak_failed))
break
}
isFirstSegment = false
val segments = TtsSegmenter.split(cleanMessage)
AppLogger.d(TAG, "speech[segments] count=${segments.size} lengths=${segments.joinToString(prefix = "[", postfix = "]") { it.length.toString() }}")
var isFirstSegment = true
for ((index, segment) in segments.withIndex()) {
if (_isSpeechPaused.value) {
logSpeechState("waitResume", "segmentIndex=$index")
isSpeechPaused.filter { !it }.first()
logSpeechState("resumeObserved", "segmentIndex=$index")
}
} catch (e: CancellationException) {
logSpeechState(
"speakMessage.cancelled",
"paused=${_isSpeechPaused.value} session=${_isSpeechSessionActive.value} message=${e.message}"

AppLogger.d(
TAG,
"speech[segmentSpeak] index=$index/${segments.lastIndex} interrupt=${if (isFirstSegment) interrupt else false} len=${segment.length} preview=\"${speechPreview(segment)}\""
)
} catch (e: Exception) {
logSpeechState("speakMessage.exception", "type=${e::class.java.simpleName} message=${e.message}")
AppLogger.e(TAG, "朗读消息失败", e)
uiStateDelegate.showToast(context.getString(R.string.chat_speak_message_failed, e.message ?: "Unknown error"))
val success = currentVoiceService.speak(
text = segment,
interrupt = if (isFirstSegment) interrupt else false,
rate = null,
pitch = null
)
AppLogger.d(TAG, "speech[segmentResult] index=$index success=$success provider=${currentVoiceService.javaClass.simpleName}")

if (!success) {
logSpeechState("segmentFailed", "index=$index")
uiStateDelegate.showToast(context.getString(R.string.chat_speak_failed))
break
}
isFirstSegment = false
}
} catch (e: CancellationException) {
logSpeechState(
"speakMessage.cancelled",
"paused=${_isSpeechPaused.value} session=${_isSpeechSessionActive.value} message=${e.message}"
)
} catch (e: Exception) {
logSpeechState("speakMessage.exception", "type=${e::class.java.simpleName} message=${e.message}")
AppLogger.e(TAG, "朗读消息失败", e)
uiStateDelegate.showToast(context.getString(R.string.chat_speak_message_failed, e.message ?: "Unknown error"))
}
}

Expand All @@ -3103,8 +3184,11 @@ class ChatViewModel(private val context: Context) : ViewModel() {
try {
logSpeechState("stopSpeaking.request")
cancelSpeechControlsHide("stopSpeaking")
// 先提升代数再清空:消费者取出的残留请求会被代数比对丢弃,积压的直接出队
speakRequestGeneration.incrementAndGet()
speechPlaybackJob?.cancel()
speechPlaybackJob = null
drainSpeakRequests("stop")
ensureActiveVoiceService()?.stop()
_isSpeechPaused.value = false
_isSpeechSessionActive.value = false
Expand Down
Loading