operating-room-monitor-server/app/services/voice_resolution.py

"""Resolve pending consumable confirmation from uploaded WAV: MinIO + Baidu ASR + parse.

本模块把语音识别流程线性化为一系列「阶段」，每个失败阶段都走 `VoiceAuditEmitter.fail`
统一写审计 + trace + 抛 `SurgeryPipelineError`。成功路径走 `emitter.success`。
"""

from __future__ import annotations

import json
from dataclasses import dataclass

from fastapi.concurrency import run_in_threadpool
from loguru import logger

from sqlalchemy.ext.asyncio import async_sessionmaker

from app.baked import pipeline as bp
from app.config import Settings
from app.database import AsyncSessionLocal
from app.repositories.voice_audits import VoiceAuditRepository
from app.services.audio_wav import WavDecodeError, wav_bytes_to_pcm16k_mono_s16le
from app.services.baidu_speech import BaiduSpeechNotConfiguredError, BaiduSpeechService
from app.services.minio_audio_storage import MinioAudioStorageService, StoredAudio
from app.services.pending_confirmation_port import PendingConfirmationStore
from app.services.voice_audit_emitter import VoiceAuditContext, VoiceAuditEmitter
from app.services.voice_confirm import (
    is_rejection_phrase,
    match_voice_choice_against_candidates,
    parse_voice_choice,
)
from app.surgery_errors import SurgeryPipelineError


@dataclass(frozen=True)
class VoiceResolveResult:
    resolved_label: str | None
    rejected: bool
    asr_text: str | None
    audio_object_key: str | None
    message: str


class VoiceConfirmationService:
    """Upload audio to MinIO, run Baidu ASR, parse choice, resolve pending queue entry."""

    def __init__(
        self,
        settings: Settings,
        sessions: PendingConfirmationStore,
        baidu: BaiduSpeechService,
        minio: MinioAudioStorageService,
        audits: VoiceAuditRepository,
        session_factory: async_sessionmaker | None = None,
        audit_emitter: VoiceAuditEmitter | None = None,
    ) -> None:
        self._s = settings
        self._sessions = sessions
        self._baidu = baidu
        self._minio = minio
        self._audits = audits
        self._session_factory: async_sessionmaker = session_factory or AsyncSessionLocal
        self._emitter = audit_emitter or VoiceAuditEmitter(
            settings=settings,
            audits=audits,
            session_factory=self._session_factory,
        )

    # ------------------------------------------------------------------
    # TTS：保持对外接口不变
    # ------------------------------------------------------------------
    def synthesize_prompt_to_mp3(self, text: str) -> bytes:
        t = (text or "").strip()
        if not t:
            raise SurgeryPipelineError("TTS_TEXT_EMPTY", "提示文本为空。")
        try:
            r = self._baidu.synthesis(
                t, "zh", 1, {"spd": 5, "pit": 5, "vol": 9, "per": 0}
            )
        except BaiduSpeechNotConfiguredError as exc:
            raise SurgeryPipelineError(
                "BAIDU_NOT_CONFIGURED",
                "服务端未配置百度语音，无法合成播报音频。",
            ) from exc
        if isinstance(r, dict):
            raise SurgeryPipelineError("TTS_ERROR", f"百度 TTS 失败: {r!r}")
        return r

    def _session_trace(self, surgery_id: str):
        def _recorder(asr_text: str | None, error: str | None) -> None:
            self._sessions.record_voice_trace(
                surgery_id, asr_text=asr_text, error=error
            )

        return _recorder

    # ------------------------------------------------------------------
    # 主入口
    # ------------------------------------------------------------------
    async def resolve_from_wav(
        self,
        *,
        surgery_id: str,
        confirmation_id: str,
        wav_bytes: bytes,
        filename: str,
        content_type: str | None,
    ) -> VoiceResolveResult:
        _ = filename  # reserved for future MIME sniff

        # 1) validate_size
        if len(wav_bytes) > bp.VOICE_UPLOAD_MAX_BYTES:
            raise await self._emitter.fail(
                source="wav",
                status="invalid_audio",
                code="VOICE_AUDIO_INVALID",
                message=(
                    f"音频大小超过限制（最大 {bp.VOICE_UPLOAD_MAX_BYTES} 字节）。"
                ),
                surgery_id=surgery_id,
                confirmation_id=confirmation_id,
                ctx=VoiceAuditContext(
                    audio_content_type=content_type,
                    audio_size_bytes=len(wav_bytes),
                ),
                # 仅对大小越界的情况按原实现不记 session trace
                record_session_trace=False,
            )

        # 2) ensure_providers_configured
        if not self._minio.configured:
            raise await self._emitter.fail(
                source="wav",
                status="minio_not_configured",
                code="MINIO_NOT_CONFIGURED",
                message="服务端未配置 MinIO，无法保存语音追溯文件。",
                surgery_id=surgery_id,
                confirmation_id=confirmation_id,
                persist_audit=False,
                record_session_trace=False,
            )
        if not self._baidu.configured:
            raise await self._emitter.fail(
                source="wav",
                status="baidu_not_configured",
                code="BAIDU_NOT_CONFIGURED",
                message="服务端未配置百度语音，无法进行语音识别。",
                surgery_id=surgery_id,
                confirmation_id=confirmation_id,
                persist_audit=False,
                record_session_trace=False,
            )

        # 3) fetch_pending
        pending = self._sessions.get_pending_confirmation_by_id(
            surgery_id, confirmation_id
        )
        if pending is None:
            raise await self._emitter.fail(
                source="wav",
                status="confirmation_not_found",
                code="CONFIRMATION_NOT_FOUND",
                message="未找到该待确认项或已处理。",
                surgery_id=surgery_id,
                confirmation_id=confirmation_id,
                persist_audit=False,
                record_session_trace=False,
            )

        option_labels = [a.strip() for a, _ in pending.options if a.strip()]
        options_snapshot = json.dumps(
            [{"label": a, "confidence": b} for a, b in pending.options],
            ensure_ascii=False,
        )
        session_trace = self._session_trace(surgery_id)

        # 4) upload_wav
        stored = await self._upload_wav(
            surgery_id=surgery_id,
            confirmation_id=confirmation_id,
            wav_bytes=wav_bytes,
            content_type=content_type,
            options_snapshot=options_snapshot,
            session_trace=session_trace,
        )

        audio_ctx = VoiceAuditContext(
            audio_object_key=stored.object_key,
            audio_content_type=content_type,
            audio_size_bytes=stored.size_bytes,
            audio_sha256=stored.sha256_hex,
        )

        # 5) decode_pcm
        pcm = await self._decode_pcm(
            surgery_id=surgery_id,
            confirmation_id=confirmation_id,
            wav_bytes=wav_bytes,
            ctx=audio_ctx,
            options_snapshot=options_snapshot,
            session_trace=session_trace,
        )

        # 6) call_asr
        asr_payload = await self._call_asr(
            surgery_id=surgery_id,
            confirmation_id=confirmation_id,
            pcm=pcm,
            ctx=audio_ctx,
            options_snapshot=options_snapshot,
            session_trace=session_trace,
        )

        # 7) extract_text
        text = await self._extract_text_from_asr(
            surgery_id=surgery_id,
            confirmation_id=confirmation_id,
            asr_payload=asr_payload,
            ctx=audio_ctx,
            options_snapshot=options_snapshot,
            session_trace=session_trace,
        )
        self._sessions.record_voice_trace(surgery_id, asr_text=text, error=None)

        # 8) parse_choice
        rejected, chosen = await self._parse_choice_or_fail(
            source="wav",
            surgery_id=surgery_id,
            confirmation_id=confirmation_id,
            text=text,
            option_labels=option_labels,
            options_snapshot=options_snapshot,
            ctx=audio_ctx,
            session_trace=session_trace,
        )

        # 9) persist_success（含 session 内的 resolve_pending_confirmation）
        await self._sessions.resolve_pending_confirmation(
            surgery_id,
            confirmation_id,
            chosen_label=chosen,
            rejected=rejected,
        )
        final_status = "rejected" if rejected else "recognized"
        await self._emitter.success(
            source="wav",
            status=final_status,
            surgery_id=surgery_id,
            confirmation_id=confirmation_id,
            ctx=audio_ctx,
            asr_text=text,
            resolved_label=chosen if not rejected else None,
            rejected=rejected,
            options_snapshot_json=options_snapshot,
        )

        if rejected:
            return VoiceResolveResult(
                resolved_label=None,
                rejected=True,
                asr_text=text,
                audio_object_key=stored.object_key,
                message="已否认全部候选，未记消耗。",
            )
        return VoiceResolveResult(
            resolved_label=chosen,
            rejected=False,
            asr_text=text,
            audio_object_key=stored.object_key,
            message="已确认并记一条消耗。",
        )

    async def resolve_from_recognized_text(
        self,
        *,
        surgery_id: str,
        confirmation_id: str,
        recognized_text: str,
    ) -> VoiceResolveResult:
        """浏览器本机识别文本，不经 MinIO/百度 ASR，解析规则与 `resolve_from_wav` 一致。"""
        pending = self._sessions.get_pending_confirmation_by_id(
            surgery_id, confirmation_id
        )
        if pending is None:
            raise await self._emitter.fail(
                source="text",
                status="confirmation_not_found",
                code="CONFIRMATION_NOT_FOUND",
                message="未找到该待确认项或已处理。",
                surgery_id=surgery_id,
                confirmation_id=confirmation_id,
                persist_audit=False,
                record_session_trace=False,
            )

        option_labels = [a.strip() for a, _ in pending.options if a.strip()]
        options_snapshot = json.dumps(
            [{"label": a, "confidence": b} for a, b in pending.options],
            ensure_ascii=False,
        )
        session_trace = self._session_trace(surgery_id)

        text = (recognized_text or "").strip()
        if not text:
            raise await self._emitter.fail(
                source="text",
                status="client_stt_empty",
                code="VOICE_TEXT_EMPTY",
                message="客户端识别文本为空",
                surgery_id=surgery_id,
                confirmation_id=confirmation_id,
                options_snapshot_json=options_snapshot,
                session_trace_recorder=lambda _a, _m: self._sessions.record_voice_trace(
                    surgery_id, asr_text=None, error="empty text"
                ),
            )

        self._sessions.record_voice_trace(surgery_id, asr_text=text, error=None)

        rejected, chosen = await self._parse_choice_or_fail(
            source="text",
            surgery_id=surgery_id,
            confirmation_id=confirmation_id,
            text=text,
            option_labels=option_labels,
            options_snapshot=options_snapshot,
            ctx=VoiceAuditContext(),
            session_trace=session_trace,
            parse_status_on_failure="client_stt_parse_failed",
            parse_message_prefix=(
                "无法从文本中匹配候选项或本台手术候选清单中的耗材名称，"
                "请重试或说「不是」否认全部。"
            ),
            parse_retry_hint_still=(
                "请输入「第一个」「第二个」等或候选项全名。"
            ),
            parse_retry_hint_exhausted=(
                "请再输入序号/全名，或说「不是」否认全部。"
            ),
        )

        await self._sessions.resolve_pending_confirmation(
            surgery_id,
            confirmation_id,
            chosen_label=chosen,
            rejected=rejected,
        )
        final_status = "rejected" if rejected else "recognized"
        await self._emitter.success(
            source="text",
            status=final_status,
            surgery_id=surgery_id,
            confirmation_id=confirmation_id,
            asr_text=text,
            resolved_label=chosen if not rejected else None,
            rejected=rejected,
            options_snapshot_json=options_snapshot,
        )

        if rejected:
            return VoiceResolveResult(
                resolved_label=None,
                rejected=True,
                asr_text=text,
                audio_object_key=None,
                message="已否认全部候选，未记消耗。",
            )
        return VoiceResolveResult(
            resolved_label=chosen,
            rejected=False,
            asr_text=text,
            audio_object_key=None,
            message="已确认并记一条消耗。",
        )

    # ------------------------------------------------------------------
    # 内部阶段
    # ------------------------------------------------------------------
    async def _upload_wav(
        self,
        *,
        surgery_id: str,
        confirmation_id: str,
        wav_bytes: bytes,
        content_type: str | None,
        options_snapshot: str,
        session_trace,
    ) -> StoredAudio:
        try:
            await run_in_threadpool(self._minio.ensure_bucket)
            return await run_in_threadpool(
                lambda: self._minio.upload_voice_wav(
                    surgery_id=surgery_id,
                    confirmation_id=confirmation_id,
                    data=wav_bytes,
                    content_type=content_type,
                )
            )
        except Exception as exc:
            logger.warning("MinIO upload failed: {}", exc)
            raise await self._emitter.fail(
                source="wav",
                status="upload_failed",
                code="MINIO_UPLOAD_FAILED",
                message=f"语音文件上传失败：{exc}",
                surgery_id=surgery_id,
                confirmation_id=confirmation_id,
                ctx=VoiceAuditContext(
                    audio_content_type=content_type,
                    audio_size_bytes=len(wav_bytes),
                ),
                options_snapshot_json=options_snapshot,
                session_trace_recorder=session_trace,
            ) from exc

    async def _decode_pcm(
        self,
        *,
        surgery_id: str,
        confirmation_id: str,
        wav_bytes: bytes,
        ctx: VoiceAuditContext,
        options_snapshot: str,
        session_trace,
    ) -> bytes:
        try:
            return await run_in_threadpool(wav_bytes_to_pcm16k_mono_s16le, wav_bytes)
        except WavDecodeError as exc:
            raise await self._emitter.fail(
                source="wav",
                status="invalid_audio",
                code="VOICE_AUDIO_INVALID",
                message=f"无法解析 WAV 音频：{exc}",
                surgery_id=surgery_id,
                confirmation_id=confirmation_id,
                ctx=ctx,
                options_snapshot_json=options_snapshot,
                session_trace_recorder=session_trace,
            ) from exc

    async def _call_asr(
        self,
        *,
        surgery_id: str,
        confirmation_id: str,
        pcm: bytes,
        ctx: VoiceAuditContext,
        options_snapshot: str,
        session_trace,
    ) -> object:
        try:
            return await run_in_threadpool(self._baidu.asr, pcm, "pcm", 16000, None)
        except BaiduSpeechNotConfiguredError as exc:
            raise await self._emitter.fail(
                source="wav",
                status="baidu_not_configured",
                code="BAIDU_NOT_CONFIGURED",
                message=str(exc),
                surgery_id=surgery_id,
                confirmation_id=confirmation_id,
                ctx=ctx,
                persist_audit=False,
                record_session_trace=False,
            ) from exc
        except Exception as exc:
            raise await self._emitter.fail(
                source="wav",
                status="asr_failed",
                code="VOICE_ASR_FAILED",
                message=f"语音识别调用失败：{exc}",
                surgery_id=surgery_id,
                confirmation_id=confirmation_id,
                ctx=ctx,
                options_snapshot_json=options_snapshot,
                session_trace_recorder=session_trace,
            ) from exc

    async def _extract_text_from_asr(
        self,
        *,
        surgery_id: str,
        confirmation_id: str,
        asr_payload: object,
        ctx: VoiceAuditContext,
        options_snapshot: str,
        session_trace,
    ) -> str:
        if not isinstance(asr_payload, dict):
            raise await self._emitter.fail(
                source="wav",
                status="asr_failed",
                code="VOICE_ASR_FAILED",
                message="ASR 返回格式异常",
                surgery_id=surgery_id,
                confirmation_id=confirmation_id,
                ctx=ctx,
                options_snapshot_json=options_snapshot,
                session_trace_recorder=session_trace,
            )
        if asr_payload.get("err_no") != 0:
            msg = (
                f"asr_err_{asr_payload.get('err_no')}: "
                f"{asr_payload.get('err_msg')}"
            )
            raise await self._emitter.fail(
                source="wav",
                status="asr_failed",
                code="VOICE_ASR_FAILED",
                message=msg,
                surgery_id=surgery_id,
                confirmation_id=confirmation_id,
                ctx=ctx,
                options_snapshot_json=options_snapshot,
                session_trace_recorder=session_trace,
            )

        results = asr_payload.get("result")
        text: str | None = None
        if isinstance(results, list) and results:
            text = str(results[0])
        elif isinstance(results, str):
            text = results
        text = (text or "").strip()
        if not text:
            raise await self._emitter.fail(
                source="wav",
                status="asr_failed",
                code="VOICE_ASR_FAILED",
                message="语音识别结果为空",
                surgery_id=surgery_id,
                confirmation_id=confirmation_id,
                ctx=ctx,
                options_snapshot_json=options_snapshot,
                session_trace_recorder=session_trace,
            )
        return text

    async def _parse_choice_or_fail(
        self,
        *,
        source,
        surgery_id: str,
        confirmation_id: str,
        text: str,
        option_labels: list[str],
        options_snapshot: str,
        ctx: VoiceAuditContext,
        session_trace,
        parse_status_on_failure: str = "parse_failed",
        parse_message_prefix: str = (
            "无法从语音中匹配候选项或本台手术候选清单中的耗材名称，"
            "请重试或说「不是」否认全部。"
        ),
        parse_retry_hint_still: str = (
            "请说「第一个」「第二个」等序号或候选项全名。"
        ),
        parse_retry_hint_exhausted: str = (
            "请再清晰地说序号/全名，或说「不是」否认全部。"
        ),
    ) -> tuple[bool, str | None]:
        rejected = is_rejection_phrase(text)
        chosen: str | None = None
        if not rejected:
            chosen = parse_voice_choice(text, option_labels)
            if chosen is None:
                candidates = self._sessions.get_surgery_candidate_consumables(surgery_id)
                chosen = match_voice_choice_against_candidates(text, candidates)

        if rejected or chosen:
            return rejected, chosen

        _, retry_remaining = await self._sessions.record_voice_parse_failure(
            surgery_id, confirmation_id
        )
        if retry_remaining > 0:
            if source == "wav":
                suffix = (
                    f" 本次未听清或未能解析，您还可重试 {retry_remaining} 次，"
                    f"{parse_retry_hint_still}"
                )
            else:
                suffix = (
                    f" 本次未能解析，您还可重试 {retry_remaining} 次，"
                    f"{parse_retry_hint_still}"
                )
        else:
            suffix = f" 本轮重试机会已用完，{parse_retry_hint_exhausted}"
        msg = parse_message_prefix + suffix

        raise await self._emitter.fail(
            source=source,
            status=parse_status_on_failure,
            code="VOICE_PARSE_FAILED",
            message=msg,
            surgery_id=surgery_id,
            confirmation_id=confirmation_id,
            ctx=ctx,
            asr_text=text,
            options_snapshot_json=options_snapshot,
            session_trace_recorder=session_trace,
            include_extra={
                "confirmation_id": confirmation_id,
                "retry_remaining": retry_remaining,
            },
        )