ddmitry/local-transcriber
1
0
Fork 0
Code Issues 6 Pull Requests Actions Packages Projects Releases Wiki Activity

feat(diarization): добавлено разделение транскрипта по говорящим

Browse Source 
Зачем:
- локальным транскриптам нужна структура реплик для конспектов и протоколов.

Что:
- добавлены пословные таймкоды для всех ASR-бэкендов и сведение с Sherpa-ONNX.
- реализованы CLI-флаги, деградация без потери ASR и speaker Markdown.
- добавлены проверяемый кеш моделей, тесты и документация.

Проверка:
- `pytest` — 283 passed, 1 skipped.
- `pyright` — 0 errors.
- Ruff и `git diff --check` — без ошибок.
- выполнены три контрольных прогона на реальных записях.
This commit is contained in:
Dmitriy Dementiev
2026-08-14 18:55:42 +03:00
parent 63ba41d906
commit 726718197f
23 changed files with 2522 additions and 187 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/test_backend_openvino.py
+140 -20
View File
@@ -11,8 +11,7 @@ from local_transcriber.backends.openvino import (
OpenVINOBackend,
_validate_model_dir,
)
from local_transcriber.types import UNKNOWN_LANGUAGE, Segment
from local_transcriber.types import UNKNOWN_LANGUAGE, Segment, Word
# === _resolve_repo ===
@@ -28,12 +27,18 @@ def test_model_catalog_contains_large_v3_turbo_profiles():
def test_resolve_repo_exact_match():
backend = OpenVINOBackend(compute_type_explicit=True)
assert backend._resolve_repo("medium", "int8") == ("OpenVINO/whisper-medium-int8-ov", "int8")
assert backend._resolve_repo("medium", "int8") == (
"OpenVINO/whisper-medium-int8-ov",
"int8",
)
def test_resolve_repo_large_v3_fp16():
backend = OpenVINOBackend(compute_type_explicit=True)
assert backend._resolve_repo("large-v3", "fp16") == ("OpenVINO/whisper-large-v3-fp16-ov", "fp16")
assert backend._resolve_repo("large-v3", "fp16") == (
"OpenVINO/whisper-large-v3-fp16-ov",
"fp16",
)
def test_resolve_repo_explicit_unsupported_pair_raises():
@@ -53,16 +58,17 @@ def test_resolve_repo_implicit_fallback():
"""Неявный compute_type: если int8 недоступен для base, fallback на fp16."""
backend = OpenVINOBackend(compute_type_explicit=False)
# base + int8 не существует, но base + fp16 есть
assert backend._resolve_repo("base", "int8") == ("OpenVINO/whisper-base-fp16-ov", "fp16")
assert backend._resolve_repo("base", "int8") == (
"OpenVINO/whisper-base-fp16-ov",
"fp16",
)
@pytest.mark.parametrize(
("model_name", "expected_compute_type"),
[("large-v3", "fp16"), ("large-v3-turbo", "int8")],
)
def test_resolve_repo_implicit_large_v3_profiles(
model_name, expected_compute_type
):
def test_resolve_repo_implicit_large_v3_profiles(model_name, expected_compute_type):
"""Неявный compute_type различает обычную и turbo-модель."""
backend = OpenVINOBackend(compute_type_explicit=False)
@@ -75,7 +81,10 @@ def test_resolve_repo_implicit_large_v3_profiles(
def test_resolve_repo_explicit_large_v3_int8_respected():
"""Явный --compute-type int8 для large-v3 → уважается."""
backend = OpenVINOBackend(compute_type_explicit=True)
assert backend._resolve_repo("large-v3", "int8") == ("OpenVINO/whisper-large-v3-int8-ov", "int8")
assert backend._resolve_repo("large-v3", "int8") == (
"OpenVINO/whisper-large-v3-int8-ov",
"int8",
)
@pytest.mark.parametrize("compute_type", ["int8", "fp16"])
@@ -107,6 +116,7 @@ def test_ensure_model_available_cache_hit(mock_download, tmp_path):
model_dir.mkdir()
(model_dir / "openvino_encoder_model.xml").write_text("<xml/>")
(model_dir / "openvino_decoder_model.xml").write_text("<xml/>")
(model_dir / "generation_config.json").write_text('{"alignment_heads": [[1, 2]]}')
mock_download.return_value = str(model_dir)
backend = OpenVINOBackend(compute_type_explicit=True)
@@ -125,6 +135,7 @@ def test_ensure_model_available_downloads(mock_download, tmp_path):
model_dir.mkdir()
(model_dir / "openvino_encoder_model.xml").write_text("<xml/>")
(model_dir / "openvino_decoder_model.xml").write_text("<xml/>")
(model_dir / "generation_config.json").write_text('{"alignment_heads": [[1, 2]]}')
mock_download.side_effect = [
LocalEntryNotFoundError("not cached"),
@@ -145,6 +156,7 @@ def test_large_v3_turbo_model_is_resolved_and_created(mock_download, tmp_path):
model_dir.mkdir()
(model_dir / "openvino_encoder_model.xml").write_text("<xml/>")
(model_dir / "openvino_decoder_model.xml").write_text("<xml/>")
(model_dir / "generation_config.json").write_text('{"alignment_heads": [[1, 2]]}')
mock_download.return_value = str(model_dir)
mock_ov = MagicMock()
@@ -157,12 +169,28 @@ def test_large_v3_turbo_model_is_resolved_and_created(mock_download, tmp_path):
"OpenVINO/whisper-large-v3-turbo-int8-ov",
local_files_only=True,
)
mock_ov.WhisperPipeline.assert_called_once_with(str(model_dir), "CPU")
mock_ov.WhisperPipeline.assert_called_once_with(
str(model_dir), "CPU", word_timestamps=True
)
# === create_model ===
def test_create_model_enables_word_timestamps():
mock_ov = MagicMock()
backend = OpenVINOBackend(ov_device="openvino-cpu")
with patch.dict("sys.modules", {"openvino_genai": mock_ov}):
backend.create_model("/path/to/model", "openvino-cpu", "int8")
mock_ov.WhisperPipeline.assert_called_once_with(
"/path/to/model",
"CPU",
word_timestamps=True,
)
def test_create_model_cpu():
mock_ov = MagicMock()
mock_pipeline = MagicMock()
@@ -172,7 +200,9 @@ def test_create_model_cpu():
with patch.dict("sys.modules", {"openvino_genai": mock_ov}):
model = backend.create_model("/path/to/model", "openvino-cpu", "int8")
mock_ov.WhisperPipeline.assert_called_once_with("/path/to/model", "CPU")
mock_ov.WhisperPipeline.assert_called_once_with(
"/path/to/model", "CPU", word_timestamps=True
)
assert model is mock_pipeline
assert backend.actual_ov_device == "CPU"
@@ -186,7 +216,9 @@ def test_create_model_gpu():
with patch.dict("sys.modules", {"openvino_genai": mock_ov}):
model = backend.create_model("/path/to/model", "openvino-gpu", "fp16")
mock_ov.WhisperPipeline.assert_called_once_with("/path/to/model", "GPU")
mock_ov.WhisperPipeline.assert_called_once_with(
"/path/to/model", "GPU", word_timestamps=True
)
assert model is mock_pipeline
assert backend.actual_ov_device == "GPU"
@@ -202,11 +234,16 @@ def test_create_model_openvino_auto_detects_gpu():
backend = OpenVINOBackend(ov_device="openvino")
with (
patch.dict("sys.modules", {"openvino_genai": mock_ov, "openvino": MagicMock(Core=mock_core)}),
patch.dict(
"sys.modules",
{"openvino_genai": mock_ov, "openvino": MagicMock(Core=mock_core)},
),
):
model = backend.create_model("/path/to/model", "openvino", "int8")
backend.create_model("/path/to/model", "openvino", "int8")
mock_ov.WhisperPipeline.assert_called_once_with("/path/to/model", "GPU")
mock_ov.WhisperPipeline.assert_called_once_with(
"/path/to/model", "GPU", word_timestamps=True
)
assert backend.actual_ov_device == "GPU"
@@ -221,11 +258,16 @@ def test_create_model_openvino_auto_falls_back_to_cpu():
backend = OpenVINOBackend(ov_device="openvino")
with (
patch.dict("sys.modules", {"openvino_genai": mock_ov, "openvino": MagicMock(Core=mock_core)}),
patch.dict(
"sys.modules",
{"openvino_genai": mock_ov, "openvino": MagicMock(Core=mock_core)},
),
):
model = backend.create_model("/path/to/model", "openvino", "int8")
backend.create_model("/path/to/model", "openvino", "int8")
mock_ov.WhisperPipeline.assert_called_once_with("/path/to/model", "CPU")
mock_ov.WhisperPipeline.assert_called_once_with(
"/path/to/model", "CPU", word_timestamps=True
)
assert backend.actual_ov_device == "CPU"
@@ -248,13 +290,19 @@ def test_transcribe_maps_chunks_to_segments():
mock_result = MagicMock()
mock_result.chunks = [chunk1, chunk2]
mock_result.words = [
MagicMock(start_ts=0.0, end_ts=3.5, word=" Привет мир"),
MagicMock(start_ts=3.5, end_ts=7.0, word=" Тестовый сегмент"),
]
mock_model.generate.return_value = mock_result
raw_audio = np.zeros(16000 * 10, dtype=np.float32) # 10 секунд
with patch("faster_whisper.decode_audio", return_value=raw_audio):
result = backend.transcribe(
mock_model, Path("test.mp3"), language="ru",
mock_model,
Path("test.mp3"),
language="ru",
)
assert len(result.segments) == 2
@@ -269,6 +317,61 @@ def test_transcribe_maps_chunks_to_segments():
assert call_kwargs.kwargs["return_timestamps"] is True
def test_transcribe_maps_word_level_timestamps():
backend = OpenVINOBackend()
mock_model = MagicMock()
raw_word = MagicMock()
raw_word.start_ts = 0.2
raw_word.end_ts = 0.8
raw_word.word = " Привет"
mock_result = MagicMock()
mock_result.chunks = []
mock_result.words = [raw_word]
mock_model.generate.return_value = mock_result
with patch(
"faster_whisper.decode_audio",
return_value=np.zeros(16_000, dtype=np.float32),
):
result = backend.transcribe(mock_model, Path("test.mp3"), language="ru")
assert result.words == [Word(start=0.2, end=0.8, text=" Привет")]
assert mock_model.generate.call_args.kwargs["word_timestamps"] is True
def test_transcribe_keeps_zero_duration_word_timestamp():
backend = OpenVINOBackend()
mock_model = MagicMock()
raw_word = MagicMock(start_ts=1.0, end_ts=1.0, word=" Слово")
mock_result = MagicMock(chunks=[], words=[raw_word])
mock_model.generate.return_value = mock_result
with patch(
"faster_whisper.decode_audio",
return_value=np.zeros(16_000, dtype=np.float32),
):
result = backend.transcribe(mock_model, Path("test.mp3"), language="ru")
assert result.words == [Word(start=1.0, end=1.0, text=" Слово")]
def test_transcribe_rejects_nonempty_result_without_word_timestamps():
backend = OpenVINOBackend()
chunk = MagicMock(start_ts=0.0, end_ts=1.0, text=" Текст")
mock_result = MagicMock(chunks=[chunk], words=None)
mock_model = MagicMock()
mock_model.generate.return_value = mock_result
with (
patch(
"faster_whisper.decode_audio",
return_value=np.zeros(16_000, dtype=np.float32),
),
pytest.raises(RuntimeError, match="пословные таймкоды"),
):
backend.transcribe(mock_model, Path("test.mp3"), language="ru")
def test_transcribe_calls_tolist():
"""raw_speech передаётся как list, не ndarray."""
backend = OpenVINOBackend()
@@ -314,6 +417,7 @@ def test_transcribe_calls_on_segment():
chunk.text = " Test"
mock_result = MagicMock()
mock_result.chunks = [chunk]
mock_result.words = [MagicMock(start_ts=0.0, end_ts=2.0, word=" Test")]
mock_model.generate.return_value = mock_result
raw_audio = np.zeros(16000, dtype=np.float32)
@@ -321,7 +425,10 @@ def test_transcribe_calls_on_segment():
with patch("faster_whisper.decode_audio", return_value=raw_audio):
backend.transcribe(
mock_model, Path("test.mp3"), language="en", on_segment=callback,
mock_model,
Path("test.mp3"),
language="en",
on_segment=callback,
)
callback.assert_called_once()
@@ -336,6 +443,7 @@ def test_transcribe_calls_on_segment():
def test_validate_model_dir_ok(tmp_path):
(tmp_path / "openvino_encoder_model.xml").write_text("<xml/>")
(tmp_path / "openvino_decoder_model.xml").write_text("<xml/>")
(tmp_path / "generation_config.json").write_text('{"alignment_heads": [[1, 2]]}')
_validate_model_dir(tmp_path) # should not raise
@@ -343,3 +451,15 @@ def test_validate_model_dir_missing(tmp_path):
(tmp_path / "openvino_encoder_model.xml").write_text("<xml/>")
with pytest.raises(ValueError, match="openvino_decoder_model.xml"):
_validate_model_dir(tmp_path)
def test_validate_model_dir_requires_alignment_heads_for_word_timestamps(tmp_path):
(tmp_path / "openvino_encoder_model.xml").write_text("<xml/>")
(tmp_path / "openvino_decoder_model.xml").write_text("<xml/>")
(tmp_path / "generation_config.json").write_text(
'{"alignment_heads": []}',
encoding="utf-8",
)
with pytest.raises(ValueError, match="alignment_heads"):
_validate_model_dir(tmp_path)