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

feat(config): добавлены device-aware дефолты и результаты тестов

Browse Source 
- Зачем:
  - тестирование на реальных записях показало, что int8 даёт галлюцинации на длинных файлах, auto-detect языка ошибается — нужны оптимальные дефолты по устройству.
- Что:
  - дефолты: medium float16 (GPU), medium float32 (CPU), language=ru.
  - добавлены DEVICE_DEFAULTS и apply_device_defaults() в config.py.
  - убран preprocessor_config.json из обязательных файлов модели (отсутствует у medium).
  - README обновлён: таблицы скоростей, качества, результаты тестирования compute_type.
- Проверка:
  - uv run pytest — 98 passed, 1 skipped.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
ddadmin
2026-03-18 21:43:06 +03:00
co-authored by Claude Opus 4.6
parent e28232ef50
commit b7e6ab634a
6 changed files with 163 additions and 70 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/test_cli.py
+26 -26
View File
@@ -43,7 +43,7 @@ def _single_patches(result=None, tmp_file=None, actual_device="cpu"):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", return_value=tmp_file),
patch("local_transcriber.cli.detect_device", return_value=actual_device),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", return_value=(model, actual_device)),
patch("local_transcriber.cli._transcribe_file", return_value=tfr),
patch("local_transcriber.cli.write_transcript"),
@@ -74,7 +74,7 @@ def test_cli_default_options_passed_to_transcribe(tmp_path):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", return_value=audio),
patch("local_transcriber.cli.detect_device", return_value="cpu"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", return_value=(model, "cpu")),
patch("local_transcriber.cli._transcribe_file", mock_transcribe_file),
patch("local_transcriber.cli.write_transcript"),
@@ -82,9 +82,9 @@ def test_cli_default_options_passed_to_transcribe(tmp_path):
runner.invoke(app, [str(audio)])
call_kwargs = mock_transcribe_file.call_args[1]
assert call_kwargs["model_name"] == "/models/large-v3"
assert call_kwargs["compute_type"] == "int8"
assert call_kwargs["language"] is None # "auto" → None
assert call_kwargs["model_name"] == "/models/medium"
assert call_kwargs["compute_type"] == "float32"
assert call_kwargs["language"] == "ru"
assert call_kwargs["on_segment"] is None # verbose=False
@@ -134,7 +134,7 @@ def test_cli_verbose_passes_on_segment_callback(tmp_path):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", return_value=audio),
patch("local_transcriber.cli.detect_device", return_value="cpu"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", return_value=(model, "cpu")),
patch("local_transcriber.cli._transcribe_file", mock_transcribe_file),
patch("local_transcriber.cli.write_transcript"),
@@ -173,7 +173,7 @@ def test_cli_default_output_path(tmp_path):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", return_value=audio),
patch("local_transcriber.cli.detect_device", return_value="cpu"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", return_value=(model, "cpu")),
patch("local_transcriber.cli._transcribe_file", return_value=tfr),
patch("local_transcriber.cli.write_transcript", mock_write),
@@ -199,7 +199,7 @@ def test_cli_custom_output_path(tmp_path):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", return_value=audio),
patch("local_transcriber.cli.detect_device", return_value="cpu"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", return_value=(model, "cpu")),
patch("local_transcriber.cli._transcribe_file", return_value=tfr),
patch("local_transcriber.cli.write_transcript", mock_write),
@@ -236,7 +236,7 @@ def test_cli_passes_status_callback_to_transcribe(tmp_path):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", return_value=audio),
patch("local_transcriber.cli.detect_device", return_value="cpu"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", return_value=(model, "cpu")),
patch("local_transcriber.cli._transcribe_file", mock_transcribe_file),
patch("local_transcriber.cli.write_transcript"),
@@ -284,7 +284,7 @@ def test_cli_windows_cuda_diagnostic(tmp_path):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", return_value=audio),
patch("local_transcriber.cli.detect_device", return_value="cuda"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", return_value=(model, "cuda")),
patch("local_transcriber.cli._transcribe_file", side_effect=RuntimeError("CUDA error: no device")),
patch("local_transcriber.cli.sys") as mock_sys,
@@ -308,7 +308,7 @@ def test_cli_linux_cuda_error_no_windows_hint(tmp_path):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", return_value=audio),
patch("local_transcriber.cli.detect_device", return_value="cuda"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", return_value=(model, "cuda")),
patch("local_transcriber.cli._transcribe_file", side_effect=RuntimeError("CUDA error: no device")),
patch("local_transcriber.cli.sys") as mock_sys,
@@ -333,7 +333,7 @@ def test_cli_device_fallback_warning(tmp_path):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", return_value=audio),
patch("local_transcriber.cli.detect_device", return_value="cuda"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", return_value=(model, "cuda")),
patch("local_transcriber.cli._transcribe_file", return_value=tfr),
patch("local_transcriber.cli.write_transcript"),
@@ -357,7 +357,7 @@ def test_cli_strict_device_passed_to_transcribe(tmp_path):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", return_value=audio),
patch("local_transcriber.cli.detect_device", return_value="cuda"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", return_value=(model, "cuda")),
patch("local_transcriber.cli._transcribe_file", mock_transcribe_file),
patch("local_transcriber.cli.write_transcript"),
@@ -377,7 +377,7 @@ def test_cli_strict_device_passed_to_transcribe(tmp_path):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", return_value=audio),
patch("local_transcriber.cli.detect_device", return_value="cpu"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", return_value=(model, "cpu")),
patch("local_transcriber.cli._transcribe_file", mock_transcribe_file),
patch("local_transcriber.cli.write_transcript"),
@@ -398,7 +398,7 @@ def test_cli_keyboard_interrupt(tmp_path):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", return_value=audio),
patch("local_transcriber.cli.detect_device", return_value="cpu"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", return_value=(model, "cpu")),
patch("local_transcriber.cli._transcribe_file", side_effect=KeyboardInterrupt),
patch("local_transcriber.cli.write_transcript"),
@@ -432,7 +432,7 @@ def test_cli_unexpected_error_verbose_traceback(tmp_path):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", return_value=audio),
patch("local_transcriber.cli.detect_device", return_value="cpu"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", return_value=(model, "cpu")),
patch("local_transcriber.cli._transcribe_file", side_effect=RuntimeError("unexpected boom")),
patch("local_transcriber.cli.write_transcript"),
@@ -454,7 +454,7 @@ def test_cli_unexpected_error_no_verbose_hint(tmp_path):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", return_value=audio),
patch("local_transcriber.cli.detect_device", return_value="cpu"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", return_value=(model, "cpu")),
patch("local_transcriber.cli._transcribe_file", side_effect=RuntimeError("unexpected boom")),
patch("local_transcriber.cli.write_transcript"),
@@ -484,7 +484,7 @@ def test_cli_batch_two_files(tmp_path):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", side_effect=lambda p: p),
patch("local_transcriber.cli.detect_device", return_value="cpu"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", return_value=(model, "cpu")),
patch("local_transcriber.cli._transcribe_file", return_value=tfr),
patch("local_transcriber.cli.write_transcript"),
@@ -512,7 +512,7 @@ def test_cli_batch_skips_existing(tmp_path):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", side_effect=lambda p: p),
patch("local_transcriber.cli.detect_device", return_value="cpu"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", return_value=(model, "cpu")),
patch("local_transcriber.cli._transcribe_file", return_value=tfr),
patch("local_transcriber.cli.write_transcript"),
@@ -563,7 +563,7 @@ def test_cli_batch_force_overwrites(tmp_path):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", side_effect=lambda p: p),
patch("local_transcriber.cli.detect_device", return_value="cpu"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", return_value=(model, "cpu")),
patch("local_transcriber.cli._transcribe_file", return_value=tfr),
patch("local_transcriber.cli.write_transcript"),
@@ -598,7 +598,7 @@ def test_cli_batch_per_file_error(tmp_path):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", side_effect=lambda p: p),
patch("local_transcriber.cli.detect_device", return_value="cpu"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", return_value=(model, "cpu")),
patch("local_transcriber.cli._transcribe_file", side_effect=transcribe_side_effect),
patch("local_transcriber.cli.write_transcript"),
@@ -630,7 +630,7 @@ def test_cli_batch_invalid_in_prescan(tmp_path):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", side_effect=validate_side_effect),
patch("local_transcriber.cli.detect_device", return_value="cpu"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", return_value=(model, "cpu")),
patch("local_transcriber.cli._transcribe_file", return_value=tfr),
patch("local_transcriber.cli.write_transcript"),
@@ -725,7 +725,7 @@ def test_cli_batch_fallback_warning(tmp_path):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", side_effect=lambda p: p),
patch("local_transcriber.cli.detect_device", return_value="cuda"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", return_value=(model, "cpu")),
patch("local_transcriber.cli._transcribe_file", return_value=tfr),
patch("local_transcriber.cli.write_transcript"),
@@ -753,7 +753,7 @@ def test_cli_batch_empty_speech_warning(tmp_path):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", side_effect=lambda p: p),
patch("local_transcriber.cli.detect_device", return_value="cpu"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", return_value=(model, "cpu")),
patch("local_transcriber.cli._transcribe_file", side_effect=[tfr_empty, tfr_ok]),
patch("local_transcriber.cli.write_transcript"),
@@ -784,7 +784,7 @@ def test_cli_batch_midstream_fallback_warning(tmp_path):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", side_effect=lambda p: p),
patch("local_transcriber.cli.detect_device", return_value="cuda"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", return_value=(model_gpu, "cuda")),
patch("local_transcriber.cli._transcribe_file", side_effect=[tfr_fallback, tfr_ok]),
patch("local_transcriber.cli.write_transcript"),
@@ -811,7 +811,7 @@ def test_cli_batch_model_loaded_once(tmp_path):
patch("local_transcriber.cli.check_ffmpeg"),
patch("local_transcriber.cli.validate_input_file", side_effect=lambda p: p),
patch("local_transcriber.cli.detect_device", return_value="cpu"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/large-v3"),
patch("local_transcriber.cli.ensure_model_available", return_value="/models/medium"),
patch("local_transcriber.cli.load_model", mock_load_model),
patch("local_transcriber.cli._transcribe_file", return_value=tfr),
patch("local_transcriber.cli.write_transcript"),
tests/test_config.py
+37 -3
View File
@@ -4,6 +4,7 @@ from unittest.mock import patch
import pytest
from local_transcriber.config import (
apply_device_defaults,
find_config_file,
load_config,
resolve_defaults,
@@ -89,8 +90,41 @@ def test_resolve_defaults_hardcoded_fallback():
{"model": None, "language": None, "device": None, "compute_type": None}, {}
)
assert result == {
"model": "large-v3",
"language": "auto",
"model": "medium",
"language": "ru",
"device": "auto",
"compute_type": "int8",
"compute_type": "float32",
}
def test_apply_device_defaults_cuda():
defaults = {"model": "medium", "language": "ru", "device": "auto", "compute_type": "float32"}
cli = {"model": None, "language": None, "device": None, "compute_type": None}
result = apply_device_defaults(defaults, "cuda", cli, {})
assert result["model"] == "medium"
assert result["compute_type"] == "float16"
def test_apply_device_defaults_cpu():
defaults = {"model": "medium", "language": "ru", "device": "auto", "compute_type": "float32"}
cli = {"model": None, "language": None, "device": None, "compute_type": None}
result = apply_device_defaults(defaults, "cpu", cli, {})
assert result["model"] == "medium"
assert result["compute_type"] == "float32"
def test_apply_device_defaults_cli_overrides():
defaults = {"model": "large-v3", "language": "ru", "device": "auto", "compute_type": "int8"}
cli = {"model": "large-v3", "language": None, "device": None, "compute_type": "int8"}
result = apply_device_defaults(defaults, "cuda", cli, {})
assert result["model"] == "large-v3"
assert result["compute_type"] == "int8"
def test_apply_device_defaults_config_overrides():
defaults = {"model": "small", "language": "ru", "device": "auto", "compute_type": "int8"}
cli = {"model": None, "language": None, "device": None, "compute_type": None}
config = {"model": "small", "compute_type": "int8"}
result = apply_device_defaults(defaults, "cuda", cli, config)
assert result["model"] == "small"
assert result["compute_type"] == "int8"