Enhance pipeline execution and voice preview handling for graceful call termination
- Introduce mechanisms in the pipeline to ensure that the end call process waits for the completion of the end speech before hanging up, improving user experience during call termination. - Update the useVoicePreview hook to handle server-initiated call endings gracefully, distinguishing between normal and error disconnections. - Adjust TTS stop frame timeout settings to optimize the timing of call terminations, ensuring timely responses without unnecessary delays. - Refactor related components to support the new end call logic, enhancing overall workflow management and user interaction.
This commit is contained in:
Xin Wang
@@ -18,6 +18,8 @@ from pipecat.adapters.schemas.function_schema import FunctionSchema
|
||||
from pipecat.adapters.schemas.tools_schema import ToolsSchema
|
||||
from pipecat.audio.vad.silero import SileroVADAnalyzer
|
||||
from pipecat.frames.frames import (
|
||||
BotStartedSpeakingFrame,
|
||||
BotStoppedSpeakingFrame,
|
||||
EndFrame,
|
||||
InputTransportMessageFrame,
|
||||
InterruptionFrame,
|
||||
@@ -222,6 +224,10 @@ async def run_pipeline(transport, cfg: AssistantConfig) -> None:
|
||||
"current": engine.start_id if workflow_active else None,
|
||||
"ended": False,
|
||||
"turns_in_node": 0,
|
||||
# 结束流程的精确计时:只在「结束节点自己的结束语」真正说完时挂断。
|
||||
"end_turn_id": None, # 结束节点回复的 turn_id(其 text_start 在 ended 之后)
|
||||
"end_armed": False, # 结束语文本已生成完(已下发 data channel)
|
||||
"end_speaking": False, # 结束语音频已开始播报
|
||||
"end_frame_queued": False,
|
||||
}
|
||||
history: list[dict] = []
|
||||
@@ -256,6 +262,35 @@ async def run_pipeline(transport, cfg: AssistantConfig) -> None:
|
||||
assistant_aggregator = PassthroughLLMAssistantAggregator(context)
|
||||
text_input = TextInputProcessor()
|
||||
|
||||
# 结束节点:等结束语「说完」(BotStoppedSpeakingFrame)再挂断,确保结束语的
|
||||
# 文字(走 data channel)与音频都已下发,避免前端只听到声音、看不到文字。
|
||||
worker_holder: dict = {}
|
||||
|
||||
class EndCallAfterSpeech(FrameProcessor):
|
||||
async def process_frame(self, frame, direction: FrameDirection):
|
||||
await super().process_frame(frame, direction)
|
||||
await self.push_frame(frame, direction)
|
||||
# 结束语文本生成完(end_armed)→ 其音频开始(end_speaking)→ 音频说完才挂断。
|
||||
# 配对 started/stopped,避免被结束节点之前的话(如先答一句再转移)的
|
||||
# stopped 事件提前触发,导致结束语被截断。
|
||||
if isinstance(frame, BotStartedSpeakingFrame) and wf_state["end_armed"]:
|
||||
wf_state["end_speaking"] = True
|
||||
elif (
|
||||
isinstance(frame, BotStoppedSpeakingFrame)
|
||||
and wf_state["end_speaking"]
|
||||
and not wf_state["end_frame_queued"]
|
||||
and worker_holder.get("worker") is not None
|
||||
):
|
||||
wf_state["end_frame_queued"] = True
|
||||
logger.info("结束语播报完毕,挂断通话")
|
||||
# 先告知前端这是正常结束(而非连接异常),再优雅挂断
|
||||
await worker_holder["worker"].queue_frame(
|
||||
OutputTransportMessageUrgentFrame(
|
||||
message={"type": "call-ended", "reason": "completed"}
|
||||
)
|
||||
)
|
||||
await worker_holder["worker"].queue_frame(EndFrame())
|
||||
|
||||
pipeline = Pipeline(
|
||||
[
|
||||
transport.input(),
|
||||
@@ -268,6 +303,7 @@ async def run_pipeline(transport, cfg: AssistantConfig) -> None:
|
||||
# waiting for a TTS provider to emit spoken-text/timestamp frames.
|
||||
assistant_aggregator,
|
||||
tts,
|
||||
EndCallAfterSpeech(),
|
||||
transport.output(),
|
||||
]
|
||||
)
|
||||
@@ -279,6 +315,7 @@ async def run_pipeline(transport, cfg: AssistantConfig) -> None:
|
||||
),
|
||||
enable_rtvi=False,
|
||||
)
|
||||
worker_holder["worker"] = worker
|
||||
|
||||
async def queue_transcript(role: str, content: str, timestamp: str) -> None:
|
||||
if content:
|
||||
@@ -357,7 +394,9 @@ async def run_pipeline(transport, cfg: AssistantConfig) -> None:
|
||||
|
||||
async def handler(params):
|
||||
logger.info(f"LLM 触发转移 → {engine.name(target)}")
|
||||
await speak_transition(edge)
|
||||
# 进结束节点不播过渡语(结束语本身就是收尾,避免打断挂断时序)
|
||||
if not engine.is_end(target):
|
||||
await speak_transition(edge)
|
||||
await go_to_node(target)
|
||||
# 返回工具结果,pipecat 随即在新节点的提示/工具下继续生成
|
||||
await params.result_callback({"status": "ok"})
|
||||
@@ -381,7 +420,8 @@ async def run_pipeline(transport, cfg: AssistantConfig) -> None:
|
||||
)
|
||||
if target and target != wf_state["current"]:
|
||||
logger.info(f"文本兜底触发转移 → {engine.name(target)}")
|
||||
await speak_transition(engine.find_edge(wf_state["current"], target))
|
||||
if not engine.is_end(target):
|
||||
await speak_transition(engine.find_edge(wf_state["current"], target))
|
||||
# 仅切换节点提示/工具,下一轮用户输入即在新节点处理
|
||||
await go_to_node(target)
|
||||
|
||||
@@ -411,6 +451,14 @@ async def run_pipeline(transport, cfg: AssistantConfig) -> None:
|
||||
|
||||
@assistant_aggregator.event_handler("on_assistant_text_start")
|
||||
async def on_assistant_text_start(_aggregator, turn_id, timestamp):
|
||||
# 进入结束节点后,第一条「开始生成」的回复就是结束节点自己的结束语
|
||||
# (其 text_start 发生在 ended 置位之后,不会误认转移前的那句)。
|
||||
if (
|
||||
workflow_active
|
||||
and wf_state["ended"]
|
||||
and wf_state["end_turn_id"] is None
|
||||
):
|
||||
wf_state["end_turn_id"] = turn_id
|
||||
await worker.queue_frame(
|
||||
OutputTransportMessageUrgentFrame(
|
||||
message={
|
||||
@@ -449,13 +497,11 @@ async def run_pipeline(transport, cfg: AssistantConfig) -> None:
|
||||
# 正常情况下转移由 LLM 直接调用转移工具完成(go_to_node),无需这里处理。
|
||||
if content and not interrupted and workflow_active:
|
||||
history.append({"role": "assistant", "content": content})
|
||||
if wf_state["ended"]:
|
||||
# 结束节点:说完结束语后挂断,不再继续多轮对话
|
||||
if not wf_state["end_frame_queued"]:
|
||||
wf_state["end_frame_queued"] = True
|
||||
logger.info("结束节点结束语已播报,挂断通话")
|
||||
await worker.queue_frame(EndFrame())
|
||||
else:
|
||||
if turn_id == wf_state["end_turn_id"]:
|
||||
# 结束节点的结束语文本已生成完(也已下发 data channel),武装挂断;
|
||||
# 真正的 EndFrame 由 EndCallAfterSpeech 在结束语「说完」时排入。
|
||||
wf_state["end_armed"] = True
|
||||
elif not wf_state["ended"]:
|
||||
wf_state["turns_in_node"] += 1
|
||||
await fallback_route()
|
||||
elif content and not interrupted:
|
||||
|
||||
Reference in New Issue
Block a user