feat(onnx-asr): реализовать transcribe с сегментацией через VAD

- Зачем:
  - Основной метод бэкенда — транскрипция аудиофайла в сегменты с временными метками.
- Что:
  - метод transcribe декодирует аудио через faster_whisper.decode_audio, затем вызывает model.recognize() с VAD-сегментацией.
  - каждый VAD-сегмент преобразуется в проектную структуру Segment.
  - поддержка колбэков on_segment, on_status.
  - написаны 4 теста: сбор сегментов, вызов on_segment, передача языка, обработка пустого аудио.
- Проверка:
  - uv run pytest tests/test_onnx_asr.py -v (14 passed)

src/local_transcriber/backends/onnx_asr.py

@@ -3,8 +3,11 @@
 from __future__ import annotations
 
 from collections.abc import Callable
+from pathlib import Path
 from typing import Any
 
+from local_transcriber.types import Segment, TranscribeResult
+
 MODEL_ALIASES: dict[str, str] = {
     "gigaam-v3": "gigaam-v3-ctc",
     "parakeet-v3": "nemo-parakeet-tdt-0.6b-v3",
@@ -62,6 +65,52 @@ class OnnxAsrBackend:
         self._vad = vad
         return model.with_vad(vad)
 
+    def transcribe(
+        self,
+        model: Any,
+        file_path: Path,
+        language: str | None,
+        on_segment: Callable[[Segment], None] | None = None,
+        on_status: Callable[[str], None] | None = None,
+    ) -> TranscribeResult:
+        """Transcribes audio file using onnx-asr model with VAD.
+
+        model: result of create_model() — a SegmentResultsAsrAdapter.
+        file_path: path to audio/video file (any format supported by faster-whisper decode).
+        language: language code (e.g. "ru", "en") — only meaningful for multilingual models.
+        """
+        from faster_whisper import decode_audio
+
+        _notify(on_status, "Загружаю аудио...")
+        audio_array = decode_audio(str(file_path), sampling_rate=16000)
+        duration = len(audio_array) / 16000.0
+
+        _notify(on_status, "Транскрибирую (onnx-asr)...")
+        segments: list[Segment] = []
+        detected_language = language or "unknown"
+
+        for vad_seg in model.recognize(audio_array, 16000, language=language):
+            seg = Segment(
+                start=max(0.0, vad_seg.start_ts),
+                end=max(0.0, vad_seg.end_ts),
+                text=vad_seg.text,
+            )
+            if on_segment is not None:
+                on_segment(seg)
+            segments.append(seg)
+            _notify(
+                on_status,
+                f"Транскрибирую (onnx-asr)... [{len(segments)} сегм.]",
+            )
+
+        return TranscribeResult(
+            segments=segments,
+            language=detected_language,
+            language_probability=1.0 if language else 0.0,
+            duration=duration,
+            device_used="",  # оркестратор проставит
+        )
+
     def _resolve_model(self, model_name: str) -> str:
         """Resolve alias to onnx-asr model name. Raw names pass through."""
         if model_name in MODEL_ALIASES:
@@ -74,3 +123,8 @@ class OnnxAsrBackend:
             f"Доступные алиасы: {SUPPORTED_ALIASES}. "
             f"Либо укажите полное имя модели onnx-asr."
         )
+
+
+def _notify(on_status: Callable[[str], None] | None, message: str) -> None:
+    if on_status is not None:
+        on_status(message)

tests/test_onnx_asr.py

@@ -1,7 +1,19 @@
 """Tests for onnx-asr backend."""
 
 import pytest
+from pathlib import Path
+
 from local_transcriber.backends.onnx_asr import OnnxAsrBackend, MODEL_ALIASES
+from local_transcriber.types import Segment, TranscribeResult
+
+
+class FakeVadSegment:
+    """Mimics onnx-asr SegmentResult."""
+
+    def __init__(self, start_ts, end_ts, text):
+        self.start_ts = start_ts
+        self.end_ts = end_ts
+        self.text = text
 
 
 class TestEnsureModelAvailable:
@@ -95,6 +107,113 @@ class TestCreateModel:
         assert calls == ["fp16"]
 
 
+class TestTranscribe:
+    def test_transcribe_collects_segments(self, monkeypatch, tmp_path):
+        """Verify transcribe maps VAD segments to project Segments."""
+        wav_file = tmp_path / "test.wav"
+        wav_file.write_bytes(b"fake audio")
+
+        audio_samples = [0.0] * 16000  # 1 second of silence
+
+        def fake_decode_audio(path, sampling_rate=16000):
+            import numpy as np
+            return np.array(audio_samples, dtype=np.float32)
+
+        class FakeModel:
+            def recognize(self, waveform, sample_rate, language=None):
+                yield FakeVadSegment(0.0, 1.0, "hello")
+                yield FakeVadSegment(1.0, 2.5, "world")
+
+        monkeypatch.setattr("faster_whisper.decode_audio", fake_decode_audio)
+
+        backend = OnnxAsrBackend()
+        backend.actual_compute_type = "int8"
+        result = backend.transcribe(
+            FakeModel(), wav_file, language=None,
+        )
+
+        assert isinstance(result, TranscribeResult)
+        assert len(result.segments) == 2
+        assert result.segments[0] == Segment(start=0.0, end=1.0, text="hello")
+        assert result.segments[1] == Segment(start=1.0, end=2.5, text="world")
+        assert result.duration == 1.0  # 16000 samples / 16000 Hz
+
+    def test_transcribe_calls_on_segment(self, monkeypatch, tmp_path):
+        """Verify on_segment callback is invoked per segment."""
+        wav_file = tmp_path / "test.wav"
+        wav_file.write_bytes(b"fake audio")
+
+        def fake_decode_audio(path, sampling_rate=16000):
+            import numpy as np
+            return np.array([0.0] * 16000, dtype=np.float32)
+
+        segments_captured = []
+
+        class FakeModel:
+            def recognize(self, waveform, sample_rate, language=None):
+                yield FakeVadSegment(0.0, 2.0, "one")
+                yield FakeVadSegment(2.0, 4.0, "two")
+
+        monkeypatch.setattr("faster_whisper.decode_audio", fake_decode_audio)
+
+        backend = OnnxAsrBackend()
+        result = backend.transcribe(
+            FakeModel(), wav_file, language=None,
+            on_segment=lambda s: segments_captured.append(s),
+        )
+
+        assert len(segments_captured) == 2
+        assert segments_captured[0].text == "one"
+        assert segments_captured[1].text == "two"
+
+    def test_transcribe_passes_language(self, monkeypatch, tmp_path):
+        """Verify language is passed to recognize()."""
+        wav_file = tmp_path / "test.wav"
+        wav_file.write_bytes(b"fake audio")
+
+        def fake_decode_audio(path, sampling_rate=16000):
+            import numpy as np
+            return np.array([0.0] * 16000, dtype=np.float32)
+
+        lang_received = []
+
+        class FakeModel:
+            def recognize(self, waveform, sample_rate, language=None):
+                lang_received.append(language)
+                yield FakeVadSegment(0.0, 1.0, "text")
+
+        monkeypatch.setattr("faster_whisper.decode_audio", fake_decode_audio)
+
+        backend = OnnxAsrBackend()
+        backend.transcribe(FakeModel(), wav_file, language="ru")
+
+        assert lang_received == ["ru"]
+
+    def test_transcribe_empty_audio(self, monkeypatch, tmp_path):
+        """Verify zero segments for silent audio."""
+        wav_file = tmp_path / "test.wav"
+        wav_file.write_bytes(b"fake audio")
+
+        def fake_decode_audio(path, sampling_rate=16000):
+            import numpy as np
+            return np.array([0.0] * 16000, dtype=np.float32)
+
+        class FakeModel:
+            def recognize(self, waveform, sample_rate, language=None):
+                # No segments yielded
+                if False:
+                    yield
+
+        monkeypatch.setattr("faster_whisper.decode_audio", fake_decode_audio)
+
+        backend = OnnxAsrBackend()
+        result = backend.transcribe(FakeModel(), wav_file, language=None)
+
+        assert len(result.segments) == 0
+        assert result.language == "unknown"
+        assert result.duration == 1.0
+
+
 class TestModelAliases:
     def test_gigaam_v3_resolves(self):
         backend = OnnxAsrBackend()