Этап 2: озвучка ответов через Silero

- Потоковая озвучка: ответ режется на фразы, синтез и воспроизведение в отдельных потоках
- Подготовка текста: числа словами, без markdown, ссылок и блоков кода
- Модель Silero скачивается в LOCALAPPDATA и прогревается при старте
- Состояние «говорю», переключатель озвучки в трее, «Стоп» прерывает речь
- Тесты нарезки текста и оркестратора озвучки

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

tests/test_speaker.py

@@ -0,0 +1,126 @@
+import threading
+import time
+from collections.abc import Callable
+
+import numpy as np
+import pytest
+from PySide6.QtCore import QCoreApplication
+
+from agr_assistent.core.speech import Speaker
+
+
+@pytest.fixture(scope="module")
+def qapp() -> QCoreApplication:
+    return QCoreApplication.instance() or QCoreApplication([])
+
+
+class FakeEngine:
+    sample_rate = 1000
+
+    def __init__(self, fail_on: str | None = None) -> None:
+        self.fail_on = fail_on
+
+    def load(self) -> None:
+        pass
+
+    def synthesize(self, text: str) -> np.ndarray:
+        if text == self.fail_on:
+            raise RuntimeError("boom")
+        return np.zeros(10, dtype=np.float32)
+
+
+class FakePlayer:
+    def __init__(self, play_seconds: float = 0.0) -> None:
+        self.play_seconds = play_seconds
+        self.played = 0
+        self.finished = 0
+        self.aborted = 0
+        self.lock = threading.Lock()
+
+    def play(self, audio: np.ndarray, sample_rate: int, should_continue: Callable[[], bool]) -> None:
+        deadline = time.monotonic() + self.play_seconds
+        while time.monotonic() < deadline:
+            if not should_continue():
+                self.abort()
+                return
+            time.sleep(0.005)
+        with self.lock:
+            self.played += 1
+
+    def finish(self) -> None:
+        self.finished += 1
+
+    def abort(self) -> None:
+        self.aborted += 1
+
+
+def _wait_until(condition: Callable[[], bool], timeout: float = 3.0) -> None:
+    deadline = time.monotonic() + timeout
+    while not condition():
+        assert time.monotonic() < deadline, "условие не выполнилось вовремя"
+        QCoreApplication.processEvents()
+        time.sleep(0.005)
+
+
+def test_reply_is_spoken_sentence_by_sentence(qapp: QCoreApplication) -> None:
+    player = FakePlayer()
+    speaker = Speaker(FakeEngine(), player, enabled=True)
+    events: list[str] = []
+    speaker.playback_started.connect(lambda: events.append("started"))
+    speaker.finished.connect(lambda: events.append("finished"))
+
+    speaker.begin()
+    speaker.feed("Раз. Два! ")
+    speaker.feed("Три")
+    speaker.end()
+
+    _wait_until(lambda: "finished" in events)
+    assert events == ["started", "finished"]
+    assert player.played == 3
+    assert player.finished == 1
+    assert not speaker.is_active
+
+
+def test_stop_interrupts_playback(qapp: QCoreApplication) -> None:
+    player = FakePlayer(play_seconds=0.5)
+    speaker = Speaker(FakeEngine(), player, enabled=True)
+    finished: list[bool] = []
+    speaker.finished.connect(lambda: finished.append(True))
+
+    speaker.begin()
+    speaker.feed("Первая фраза. Вторая фраза. ")
+    _wait_until(lambda: speaker.is_playing)
+    speaker.stop()
+
+    assert finished == [True]
+    _wait_until(lambda: player.aborted >= 1)
+    time.sleep(0.1)
+    assert player.played == 0
+
+
+def test_disabled_speaker_ignores_replies(qapp: QCoreApplication) -> None:
+    player = FakePlayer()
+    speaker = Speaker(FakeEngine(), player, enabled=False)
+
+    speaker.begin()
+    speaker.feed("Привет. ")
+    speaker.end()
+
+    assert not speaker.is_active
+    time.sleep(0.1)
+    assert player.played == 0
+
+
+def test_synthesis_error_stops_reply_and_reports(qapp: QCoreApplication) -> None:
+    speaker = Speaker(FakeEngine(fail_on="Плохо."), FakePlayer(), enabled=True)
+    errors: list[str] = []
+    speaker.error_occurred.connect(errors.append)
+
+    speaker.begin()
+    speaker.feed("Плохо. Хорошо. Ещё. ")
+    speaker.end()
+
+    _wait_until(lambda: bool(errors))
+    assert len(errors) == 1
+    assert "boom" in errors[0]
+    assert not speaker.is_active

tests/test_speech_text.py

@@ -0,0 +1,50 @@
+from agr_assistent.tts.text import SpeechTextStream, normalize_for_speech, split_long_text
+
+
+def _feed_all(chunks: list[str]) -> list[str]:
+    stream = SpeechTextStream()
+    result = []
+    for chunk in chunks:
+        result.extend(stream.feed(chunk))
+    return result + stream.flush()
+
+
+def test_sentences_are_emitted_as_soon_as_complete() -> None:
+    stream = SpeechTextStream()
+
+    assert stream.feed("Привет! Как ") == ["Привет!"]
+    assert stream.feed("дела? Хоро") == ["Как дела?"]
+    assert stream.flush() == ["Хоро"]
+
+
+def test_decimal_numbers_and_list_markers_do_not_split() -> None:
+    assert _feed_all(["Будет 3.", "5 градуса.\n", "1. Первый пункт"]) == [
+        "Будет три целых пять десятых градуса.",
+        "Первый пункт",
+    ]
+
+
+def test_code_blocks_are_skipped() -> None:
+    chunks = ["Пример:\n``", "`python\nprint('Привет')\n", "```\nГотово."]
+
+    assert _feed_all(chunks) == ["Пример:", "Готово."]
+
+
+def test_normalize_removes_markdown_links_and_latin() -> None:
+    text = "**Важно**: смотри [документацию](https://x.io) и https://y.io, Python тоже 😀"
+
+    assert normalize_for_speech(text) == "Важно: смотри документацию и ссылка, тоже"
+
+
+def test_text_without_cyrillic_is_not_spoken() -> None:
+    assert normalize_for_speech("Hello, world! 👋") == ""
+    assert _feed_all(["Hello world.\n"]) == []
+
+
+def test_long_text_is_split_on_commas_and_spaces() -> None:
+    text = ", ".join(["слово"] * 100)
+
+    parts = split_long_text(text, limit=50)
+
+    assert all(len(part) <= 50 for part in parts)
+    assert " ".join(parts).replace(" ,", ",") == text