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

from __future__ import annotations

import asyncio
import os
import platform
import re
import shutil
import subprocess
import tempfile
from dataclasses import dataclass

from fastapi.concurrency import run_in_threadpool
from loguru import logger

from app.config import Settings
from app.services.baidu_speech import BaiduSpeechNotConfiguredError, BaiduSpeechService


_CN_DIGITS = {
    "零": 0,
    "一": 1,
    "二": 2,
    "两": 2,
    "三": 3,
    "四": 4,
    "五": 5,
    "六": 6,
    "七": 7,
    "八": 8,
    "九": 9,
    "十": 10,
}


def parse_voice_choice(asr_text: str, options: list[str]) -> str | None:
    """
    从识别文本中解析医生选择的耗材名称。
    支持：完全匹配、子串匹配、第 N 个（1/一/第一个）。
    """
    raw = (asr_text or "").strip()
    if not raw:
        return None
    normalized = raw.replace(" ", "").lower()

    for opt in options:
        if opt and opt in raw:
            return opt

    m_num = re.search(r"(\d+)", raw)
    if m_num:
        idx = int(m_num.group(1)) - 1
        if 0 <= idx < len(options):
            return options[idx]

    m_cn = re.search(r"第([一二两三四五六七八九十\d]+)个", raw)
    if m_cn:
        token = m_cn.group(1)
        if token.isdigit():
            idx = int(token) - 1
        elif token in _CN_DIGITS:
            idx = _CN_DIGITS[token] - 1
        else:
            idx = -1
        if 0 <= idx < len(options):
            return options[idx]

    for i, opt in enumerate(options):
        if not opt:
            continue
        aliases = [f"第{i + 1}个", f"第{i + 1}", f"{i + 1}号"]
        if any(a in normalized for a in aliases):
            return opt

    negatives = ("不是", "没有", "否", "无", "错")
    if any(n in raw for n in negatives):
        return None

    return None


def is_rejection_phrase(asr_text: str) -> bool:
    """医生明确否认全部候选时返回 True（须在 parse_voice_choice 之前调用）。"""
    raw = (asr_text or "").strip()
    if not raw:
        return False
    negatives = ("不是", "没有", "否", "无", "错")
    return any(n in raw for n in negatives)


def build_prompt_text(options: list[tuple[str, float]]) -> str:
    parts = ["请确认刚才使用的耗材是下面哪一项，可以说序号或名称。"]
    for i, (name, _conf) in enumerate(options, start=1):
        parts.append(f"第{i}个，{name}。")
    parts.append("若都不是请说不是。")
    return "".join(parts)


@dataclass
class VoiceAttemptResult:
    chosen_label: str | None
    asr_text: str | None
    error: str | None


class VoiceConfirmationOrchestrator:
    """服务端 TTS 播报 + ffmpeg 采集 + 百度 ASR + 文本解析。"""

    def __init__(self, settings: Settings, baidu: BaiduSpeechService) -> None:
        self._s = settings
        self._baidu = baidu
        self._lock = asyncio.Lock()

    def _ffplay_path(self) -> str | None:
        return shutil.which("ffplay")

    def _ffmpeg_path(self) -> str | None:
        return shutil.which("ffmpeg")

    def _record_pcm_ffmpeg(self, seconds: float) -> tuple[bytes | None, str | None]:
        ffmpeg = self._ffmpeg_path()
        if not ffmpeg:
            return None, "ffmpeg not found in PATH"
        system = platform.system()
        if system == "Darwin":
            dev = self._s.voice_ffmpeg_input.strip() or ":0"
            input_args = ["-f", "avfoundation", "-i", dev]
        else:
            dev = self._s.voice_ffmpeg_input.strip() or "default"
            input_args = ["-f", "alsa", "-i", dev]

        cmd = [
            ffmpeg,
            "-nostdin",
            "-loglevel",
            "error",
            "-y",
            *input_args,
            "-t",
            str(seconds),
            "-ar",
            "16000",
            "-ac",
            "1",
            "-f",
            "s16le",
            "-acodec",
            "pcm_s16le",
            "pipe:1",
        ]
        try:
            proc = subprocess.run(
                cmd,
                capture_output=True,
                timeout=seconds + 5.0,
                check=False,
            )
        except subprocess.TimeoutExpired:
            return None, "ffmpeg record timeout"
        if proc.returncode != 0:
            err = (proc.stderr or b"").decode("utf-8", errors="replace")
            return None, f"ffmpeg failed: {err or proc.returncode}"
        return proc.stdout, None

    def _play_mp3_file(self, path: str) -> str | None:
        ffplay = self._ffplay_path()
        if not ffplay:
            return "ffplay not found in PATH"
        try:
            proc = subprocess.run(
                [
                    ffplay,
                    "-nodisp",
                    "-autoexit",
                    "-loglevel",
                    "quiet",
                    path,
                ],
                capture_output=True,
                timeout=120.0,
                check=False,
            )
        except subprocess.TimeoutExpired:
            return "ffplay timeout"
        if proc.returncode != 0:
            return f"ffplay exit {proc.returncode}"
        return None

    def _synthesize_to_temp_mp3(self, text: str) -> tuple[str | None, str | None]:
        try:
            audio = self._baidu.synthesis(
                text,
                "zh",
                1,
                {"spd": 5, "pit": 5, "vol": 9, "per": 0},
            )
        except BaiduSpeechNotConfiguredError as exc:
            return None, str(exc)
        if isinstance(audio, dict):
            return None, f"TTS error: {audio!r}"
        tmp = tempfile.NamedTemporaryFile(suffix=".mp3", delete=False)
        try:
            tmp.write(audio)
            tmp.flush()
            path = tmp.name
        finally:
            tmp.close()
        return path, None

    async def run_confirmation(
        self,
        *,
        surgery_id: str,
        options: list[tuple[str, float]],
    ) -> VoiceAttemptResult:
        if not self._s.voice_confirmation_enabled:
            return VoiceAttemptResult(None, None, "voice_confirmation_disabled")
        if not options:
            return VoiceAttemptResult(None, None, "no_options")
        if not self._baidu.configured:
            return VoiceAttemptResult(None, None, "baidu_speech_not_configured")

        labels = [o[0] for o in options]
        prompt = build_prompt_text(options)
        logger.info("Voice confirm surgery={} prompt_len={}", surgery_id, len(prompt))

        async with self._lock:
            mp3_path, err = await run_in_threadpool(self._synthesize_to_temp_mp3, prompt)
            if err or not mp3_path:
                return VoiceAttemptResult(None, None, err or "tts_failed")
            try:
                play_err = await run_in_threadpool(self._play_mp3_file, mp3_path)
                if play_err:
                    return VoiceAttemptResult(None, None, play_err)
            finally:
                try:
                    os.unlink(mp3_path)
                except OSError:
                    pass

            pcm, rec_err = await run_in_threadpool(
                self._record_pcm_ffmpeg, float(self._s.voice_record_seconds)
            )
            if rec_err or not pcm:
                return VoiceAttemptResult(None, None, rec_err or "empty_audio")

        asr_payload = await run_in_threadpool(self._baidu.asr, pcm, "pcm", 16000, None)
        if not isinstance(asr_payload, dict):
            return VoiceAttemptResult(None, None, "asr_invalid_response")
        if asr_payload.get("err_no") != 0:
            return VoiceAttemptResult(
                None,
                None,
                f"asr_err_{asr_payload.get('err_no')}: {asr_payload.get('err_msg')}",
            )
        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
        if not text:
            return VoiceAttemptResult(None, None, "asr_empty_text")

        chosen = parse_voice_choice(text, labels)
        return VoiceAttemptResult(chosen, text, None)