Complete DDS training workflow and delivery package

backend/app/voice/recording.py

@@ -0,0 +1,81 @@
+"""Локальная WAV-запись обеих сторон учебного голосового вызова."""
+
+import os
+import time
+import wave
+from dataclasses import dataclass
+from pathlib import Path
+from uuid import UUID
+
+import numpy as np
+
+from app.config import get_settings
+
+TARGET_RATE = 16_000
+
+
+def recording_path(session_id: UUID) -> Path:
+    return Path(get_settings().recordings_dir).resolve() / f"{session_id}.wav"
+
+
+@dataclass(frozen=True)
+class _Segment:
+    offset: int
+    samples: np.ndarray
+
+
+class CallRecorder:
+    """Смешивает PCM16 разных частот на монотаймлайн 16 кГц.
+
+    Вход курсанта приходит по 16 кГц, TTS звонящего — по 24 кГц. Метка
+    monotonic сохраняет паузы и взаимное расположение реплик; системные часы
+    и изменение времени на хосте на запись не влияют.
+    """
+
+    def __init__(self, path: Path, *, clock=time.monotonic) -> None:
+        self.path = path
+        self._clock = clock
+        self._started = clock()
+        self._segments: list[_Segment] = []
+        self._finalized = False
+
+    def add_pcm(self, pcm: bytes, *, sample_rate: int) -> None:
+        if self._finalized or not pcm or sample_rate <= 0 or len(pcm) % 2:
+            return
+        source = np.frombuffer(pcm, dtype="<i2").astype(np.int32)
+        if source.size == 0:
+            return
+        if sample_rate != TARGET_RATE:
+            length = max(1, round(source.size * TARGET_RATE / sample_rate))
+            points = np.linspace(0, source.size - 1, length)
+            source = np.rint(np.interp(points, np.arange(source.size), source)).astype(np.int32)
+        offset = max(0, round((self._clock() - self._started) * TARGET_RATE))
+        self._segments.append(_Segment(offset=offset, samples=source))
+
+    def finalize(self) -> Path | None:
+        if self._finalized:
+            return self.path if self.path.is_file() else None
+        self._finalized = True
+        if not self._segments:
+            return None
+        total = max(item.offset + item.samples.size for item in self._segments)
+        mixed = np.zeros(total, dtype=np.int32)
+        for item in self._segments:
+            mixed[item.offset:item.offset + item.samples.size] += item.samples
+        pcm = np.clip(mixed, -32768, 32767).astype("<i2").tobytes()
+
+        self.path.parent.mkdir(parents=True, exist_ok=True)
+        temporary = self.path.with_suffix(".wav.tmp")
+        with wave.open(str(temporary), "wb") as target:
+            target.setnchannels(1)
+            target.setsampwidth(2)
+            target.setframerate(TARGET_RATE)
+            target.writeframes(pcm)
+        os.replace(temporary, self.path)
+        return self.path
+
+
+def start_recording(session_id: UUID) -> CallRecorder | None:
+    if not get_settings().record_calls:
+        return None
+    return CallRecorder(recording_path(session_id))