From: snickersun Date: Fri, 5 Jun 2026 07:58:26 +0000 (+0800) Subject: 修复静音20秒不关闭ws的问题。 X-Git-Url: http://git.snicker.com.cn/css/static/gitweb.css?a=commitdiff_plain;h=65a06a500b30578a4223860c752d8aae3cb8fbb1;p=mirrors%2FASR1603_G5.git 修复静音20秒不关闭ws的问题。 --- diff --git a/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/default_config.py b/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/default_config.py index 52557e27..9605280f 100644 --- a/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/default_config.py +++ b/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/default_config.py @@ -189,6 +189,14 @@ VC_WELCOME_SERVER_ENABLED = 0 # 协议 failed / websocket_error 是否 synthesis 播报错误原文(联调=1,量产=0) VC_ERROR_TTS_ENABLED = 1 +# 是否累积并打印 AI 回复文本(llm.delta):联调=1 可在日志看到回复;量产=0 省 recv 线程开销 +VC_LLM_DELTA_LOG = 0 +# V3 start.output_mode:audio=仅下发 TTS 音频(默认);both/text=服务端额外下发 llm.delta 文本, +# 配合 VC_LLM_DELTA_LOG=1 可在日志看到 AI 回复文字。需云端 agent 支持文本输出方有效。 +VC_OUTPUT_MODE = "audio" +# 服务端 tts.done 后立刻 clean-close WS 时,Close 前等待本地 PCM 放完尾音的最大时长(ms), +# 防"对话播报一半就停";仅对正常收尾且仍在播放时生效 +VC_CLEAN_CLOSE_DRAIN_MS = 3000 # 长对话 L1:轮次间隙无 TTS 下行累计达此秒 Close(听/想/播/录、保活帧不计;最小 5) VC_IDLE_AUTO_CLOSE_SEC = 20 diff --git a/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/products/configs/k8v3/default_config.py b/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/products/configs/k8v3/default_config.py index 035d90fa..9605280f 100644 --- a/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/products/configs/k8v3/default_config.py +++ b/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/products/configs/k8v3/default_config.py @@ -189,6 +189,11 @@ VC_WELCOME_SERVER_ENABLED = 0 # 协议 failed / websocket_error 是否 synthesis 播报错误原文(联调=1,量产=0) VC_ERROR_TTS_ENABLED = 1 +# 是否累积并打印 AI 回复文本(llm.delta):联调=1 可在日志看到回复;量产=0 省 recv 线程开销 +VC_LLM_DELTA_LOG = 0 +# V3 start.output_mode:audio=仅下发 TTS 音频(默认);both/text=服务端额外下发 llm.delta 文本, +# 配合 VC_LLM_DELTA_LOG=1 可在日志看到 AI 回复文字。需云端 agent 支持文本输出方有效。 +VC_OUTPUT_MODE = "audio" # 服务端 tts.done 后立刻 clean-close WS 时,Close 前等待本地 PCM 放完尾音的最大时长(ms), # 防"对话播报一半就停";仅对正常收尾且仍在播放时生效 VC_CLEAN_CLOSE_DRAIN_MS = 3000 diff --git a/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/vc_fsm.py b/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/vc_fsm.py index 3eb0a20c..d934e4b4 100644 --- a/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/vc_fsm.py +++ b/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/vc_fsm.py @@ -248,6 +248,21 @@ class VCStateMachine: self.mark_tts_rx_idle() self.log("local.audio_played") + def on_local_no_speech_idle(self): + # 场景①:长对话开麦后本轮无人说话,停录回 idle(会话保持、不提交)。 + # 必须把 turn 从 Listening 拉回 Idle:否则 _vc_long_idle_timer_armed() 因 + # turn==Listening 恒为 False,L1 空闲自动关无法累计,会话只能干等云端 + # IdleTimeout 报错并被 TTS 念出来。仅在未进入思考/播放链路时下拉,避免误伤。 + if self.session == VCSessionState.Started: + if self.turn in ( + VCTurnState.Listening, + VCTurnState.SpeechStreaming, + VCTurnState.SpeechEnded, + ): + self.turn = VCTurnState.Idle + self.reset_rx_phases() + self.log("local.no_speech_idle") + def on_local_play_start(self): if self.session == VCSessionState.Started: self.turn = VCTurnState.Playing diff --git a/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/vc_proto.py b/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/vc_proto.py index 5a85b2d0..6bc57347 100644 --- a/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/vc_proto.py +++ b/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/vc_proto.py @@ -186,11 +186,17 @@ def _convert_start_payload_v25_to_v3(payload): out_cfg, out_extra = _pick_v3_audio_cfg(outb.get("cfg") or {}, True) in_cfg, out_cfg = _apply_v3_opus_format(in_cfg, out_cfg) + # 输出模态可配:VC_OUTPUT_MODE 缺省 audio(仅下发 TTS 音频)。 + # 配成 both/text 时服务端才会下发 llm.delta 文本,端侧才能打印 AI 回复文字。 + out_mode = str(cfg_attr("VC_OUTPUT_MODE", "audio") or "audio").lower().strip() + if out_mode not in ("audio", "text", "both"): + out_mode = "audio" + data = { "device_id": payload.get("device_id") or "", "session_mode": session_mode, "input_mode": "audio", - "output_mode": "audio", + "output_mode": out_mode, "audio_config": { "input": in_cfg, "output": out_cfg, diff --git a/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/vc_session_io.py b/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/vc_session_io.py index 2877828e..5a2b7cfe 100644 --- a/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/vc_session_io.py +++ b/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/vc_session_io.py @@ -56,6 +56,16 @@ class VCSessionIoMixin: return None return s + def _vc_accum_llm_delta(self, raw): + """VC_LLM_DELTA_LOG=1 调试:累积 llm.delta 文本片段,llm.completed 时整段打印 AI 回复。""" + try: + d = vc_proto_normalize_server(ujson.loads(raw)) + frag = (d.get("payload") or {}).get("content", "") or "" + if frag: + self._vc_llm_reply_buf = getattr(self, "_vc_llm_reply_buf", "") + frag + except Exception: + pass + def _vc_classify_llm_tts_kind(self, payload_dialog_id=""): """设计稿 llm tts:V3 下行常无 dialog_id,用轮次状态补判为 dialog。""" if payload_dialog_id: @@ -833,6 +843,9 @@ class VCSessionIoMixin: # recv 线程会被拖住 2~3s,导致其后的 tts.started 晚到、播放迟迟不开始。 # 这里在解析前用子串快速识别并跳过(idle 计时已在解析前刷新,安全)。 if self.is_VAD != 0 and ('"llm.delta"' in recv_data): + # VC_LLM_DELTA_LOG=1 调试:累积回复文本,llm.completed 整段打印 + if cfg_bool("VC_LLM_DELTA_LOG", False): + self._vc_accum_llm_delta(recv_data) continue jsload_data = ujson.loads(recv_data) try: @@ -1089,6 +1102,7 @@ class VCSessionIoMixin: self._uplink_mark_end("vad.silence") elif jsload_data["event"] == "llm.created" :#思考(LLM) self._turn_has_llm_created = 1 + self._vc_llm_reply_buf = "" if ( self._vc_is_long_dialogue_mode() and getattr(self, "VC_record_state_flag", 0) != 0 @@ -1109,6 +1123,13 @@ class VCSessionIoMixin: ): self._vc_expect_dialog_tts = 1 elif jsload_data["event"] == "llm.completed" :#llm完成(可能无 tts) + try: + _reply = getattr(self, "_vc_llm_reply_buf", "") + if _reply: + print("[VC][LLM] reply={}".format(_reply)) + self._vc_llm_reply_buf = "" + except Exception: + pass try: self._llm_completed_ms = utime.ticks_ms() except Exception: @@ -1121,6 +1142,13 @@ class VCSessionIoMixin: if hasattr(self.fsm, "log"): self.fsm.log("server.llm.completed") elif jsload_data["event"] == "llm.delta" :#llm返回数据 + if cfg_bool("VC_LLM_DELTA_LOG", False): + try: + _frag = (jsload_data.get("payload") or {}).get("content", "") or "" + if _frag: + self._vc_llm_reply_buf = getattr(self, "_vc_llm_reply_buf", "") + _frag + except Exception: + pass if self.is_VAD == 0 and self.record_data_save_flag == 0: if self.VC_record_state_flag == 1:#llm返回数据还没停止,就先停止 self.request_uplink_stop("llm.delta") diff --git a/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/vc_session_media.py b/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/vc_session_media.py index ba360c38..d4606ac4 100644 --- a/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/vc_session_media.py +++ b/2.sdk/project/7PIN_1_5/sdk/services/microPython/ports/quectel/core/vc_session_media.py @@ -1268,6 +1268,12 @@ class VCSessionMediaMixin: if getattr(self, "_vc_no_speech_idle", 0) == 1: # 场景①:开麦后一直无有效语音,停录回 idle 省电(会话保持,不提交) self._vc_publish_led("idle", "VC_Speech_End/no_speech") + # 同步把 FSM 拉回 Idle,使 L1 空闲自动关(close_timerout)能开始累计; + # 否则 turn 卡在 Listening,会话只能等云端 IdleTimeout 报错 + try: + self.fsm.on_local_no_speech_idle() + except Exception: + pass elif ( getattr(self, "_vc_session_started", 0) == 1 and getattr(self, "Close_flag", 0) == 0