digiops/call-center
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(voice): harden noisy streaming turns

Browse Source
This commit is contained in:
Yera All
2026-04-17 00:30:41 +05:00
parent 8ca5f23e93
commit 39096e114f
6 changed files with 267 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files
services/ai_orchestrator_service/voice.py
+33
View File
@@ -930,6 +930,19 @@ def _voice_topic_prompt(language: str, caller_texts: list[str]) -> str | None:
return None
def _voice_summary_slot_prompt(language: str, context_summary: str | dict[str, Any] | None) -> str | None:
summary = load_context_summary(context_summary)
intent = str(summary.get("active_intent") or "").strip()
facts = summary.get("confirmed_facts") if isinstance(summary.get("confirmed_facts"), dict) else {}
city = str(facts.get("city") or "").strip()
branch_hint = str(facts.get("branch_hint") or "").strip()
if intent == "schedule" and city and not branch_hint:
if language == "kz":
return f"{city} qalasy boiynsha qaysy filial nemese mekenjai qyzyqtyratynyn aitnyz."
return f"По городу {city} уточните, пожалуйста, филиал или адрес."
return None
def _voice_generic_prompt(language: str) -> str:
if language == "kz":
return "Jyldam komektesu ushin eki ush sozben ne kerek ekenin aitnyz: jumys uaqyty, otinish statusy, tarif nemese operator."
@@ -1850,6 +1863,26 @@ def _voice_decision(
decision["metadata"] = v2_metadata
return decision
summary_slot_prompt = _voice_summary_slot_prompt(language, context_summary)
if not kb_results and summary_slot_prompt:
decision = {
"language": language,
"intent": "clarification",
"reply_text": summary_slot_prompt,
"confidence": 0.72,
"needs_handoff": False,
"handoff_reason": None,
"case_action": "keep_open",
"kb_refs": [],
"summary_text": "AI продолжил активный сценарий из conversation summary и уточнил недостающий слот.",
"model": "voice_policy_context_summary",
"latency_ms": 1,
}
if v2_metadata:
decision["reply_text"] = _voice_compact_reply_text(decision["reply_text"], language=language)
decision["metadata"] = {**v2_metadata, "reply_phase": "final"}
return decision
llm_decision = _voice_llm_decision(
language=language,
customer=customer,
services/ai_voice_runtime_service/media_runtime.py
+51 -15
View File
@@ -5,6 +5,7 @@ import audioop
import contextlib
import hashlib
import logging
import os
import threading
import time
import uuid
@@ -186,6 +187,10 @@ class AudioSocketMediaRuntime:
self._immediate_ack_min_ms = 700
self._v2_ack_post_gap_seconds = 0.10
self._partial_poll_interval_seconds = 0.20
self._streaming_asr_partial_poll_enabled = (
str(os.getenv("AI_VOICE_V2_STREAMING_ASR_PARTIAL_POLL_ENABLED", "0")).strip().lower()
in {"1", "true", "yes", "on"}
)
self._stable_partial_hold_seconds = 0.40
self._barge_in_trigger_ms = 220
self._streaming_asr_reopen_backoff_seconds = 2.0
@@ -197,7 +202,10 @@ class AudioSocketMediaRuntime:
@staticmethod
def _normalize_intent_text(text: str) -> str:
return " ".join(str(text or "").strip().lower().split())
compact = str(text or "").strip().lower()
for char in ",.!?;:…\"'()[]{}":
compact = compact.replace(char, " ")
return " ".join(compact.split())
@classmethod
def _is_low_signal_transcript(cls, text: str | None) -> bool:
@@ -220,9 +228,12 @@ class AudioSocketMediaRuntime:
"привет",
"слышу",
"слышно",
"твой",
"угу",
"хорошо",
"ясно",
"давай",
"поргай что это",
}
def _detect_early_intent(self, text: str) -> str:
@@ -270,7 +281,7 @@ class AudioSocketMediaRuntime:
return "Сейчас сориентирую."
if ack_kind == "clarify":
return "Сейчас уточню."
return "Сейчас подскажу."
return "Секунду."
@staticmethod
def _should_use_emotive_ack(registration: MediaRegistration, language: str | None) -> bool:
@@ -323,11 +334,10 @@ class AudioSocketMediaRuntime:
"Ага, сейчас сориентирую.",
)
return (
"Угу, сейчас подскажу.",
"Мхм, секунду.",
"Ага, сейчас подскажу.",
"Секунду.",
"Хм, секунду.",
"Хорошо, сейчас подскажу.",
"Хорошо, секунду.",
)
def _select_ack_payload(
@@ -623,6 +633,36 @@ class AudioSocketMediaRuntime:
intent = self._detect_early_intent(transcript_text)
self._update_stable_partial_intent(actor, intent)
async def _run_streaming_partial_poll(self, actor: MediaActor, utterance_generation: int) -> None:
try:
if utterance_generation != actor.utterance_generation:
return
await self._poll_streaming_partial(actor)
except asyncio.CancelledError:
raise
except Exception as exc:
logger.warning(
"audiosocket.streaming_asr_partial_failed session_id=%s generation=%s error=%s",
actor.registration.voice_session_id,
utterance_generation,
str(exc)[:500],
)
def _maybe_schedule_streaming_partial_poll(self, actor: MediaActor) -> None:
if not self._streaming_asr_partial_poll_enabled:
return
if actor.closed or not actor.asr_streaming_enabled or not actor.asr_stream_id:
return
if time.monotonic() < actor.asr_poll_due_monotonic:
return
task = actor.partial_asr_task
if task is not None and not task.done():
return
actor.partial_asr_task = asyncio.create_task(
self._run_streaming_partial_poll(actor, actor.utterance_generation),
name=f"streaming-asr-partial-{actor.registration.voice_session_id}",
)
async def _finalize_streaming_transcription(self, actor: MediaActor) -> ASRTranscription:
if not actor.asr_streaming_enabled or not actor.asr_stream_id:
raise StreamingASRUnavailable("Streaming ASR stream is not active")
@@ -756,10 +796,11 @@ class AudioSocketMediaRuntime:
language: str | None,
metadata: dict[str, Any],
ack_source: str,
ack_kind: str | None = None,
) -> None:
if actor.closed or actor.early_ack_started:
return
ack_kind = self._ack_kind_for_intent(actor.partial_intent or "unknown")
ack_kind = ack_kind or self._ack_kind_for_intent(actor.partial_intent or "unknown")
ack_text, style_hints, ack_variant = self._select_ack_payload(
actor,
language=language,
@@ -774,7 +815,7 @@ class AudioSocketMediaRuntime:
metadata={
**metadata,
"partial_transcript": actor.partial_transcript,
"early_intent": actor.partial_intent,
"early_intent": metadata.get("early_intent") or actor.partial_intent,
"ack_kind": ack_kind,
"ack_variant": ack_variant,
"voice_style": "emotive_ack" if style_hints else "neutral_ack",
@@ -1127,6 +1168,7 @@ class AudioSocketMediaRuntime:
stream_input_active = actor.input_active or actor.barge_in_pending
if actor.asr_streaming_enabled and actor.asr_stream_id and stream_input_active:
self._queue_streaming_asr_pcm(actor, pcm_frame)
self._maybe_schedule_streaming_partial_poll(actor)
elif actor.registration.voice_v2_partial_asr:
self._maybe_schedule_partial_asr(actor)
if vad_result.utterance_pcm:
@@ -1183,14 +1225,7 @@ class AudioSocketMediaRuntime:
"early_intent": partial_intent,
}
if self._should_use_voice_v2(actor.registration) and not actor.early_ack_started:
if actor.asr_streaming_enabled:
with contextlib.suppress(StreamingASRUnavailable):
await self._poll_streaming_partial(actor)
partial_transcript = str(actor.partial_transcript or "").strip()
partial_intent = str(actor.stable_partial_intent or actor.partial_intent or "").strip() or self._detect_early_intent(partial_transcript)
base_metadata["partial_transcript"] = partial_transcript
base_metadata["early_intent"] = partial_intent
else:
if not actor.asr_streaming_enabled:
partial_task = actor.partial_asr_task
if partial_task is not None and not partial_task.done():
with contextlib.suppress(asyncio.TimeoutError, asyncio.CancelledError):
@@ -1212,6 +1247,7 @@ class AudioSocketMediaRuntime:
language=actor.registration.language,
metadata=base_metadata,
ack_source="immediate_turn_close",
ack_kind="unknown",
)
if actor.asr_streaming_enabled and not actor.asr_streaming_failed:
services/shared/ai_context_summary.py
+1
View File
@@ -33,6 +33,7 @@ _LOW_SIGNAL_TEXTS = {
_CITY_ALIASES = {
"алма ата": "Алма-Ата",
"алмат": "Алмата",
"матта": "Алмата",
"астан": "Астана",
"актау": "Актау",
"актоб": "Актобе",